├── .gitignore ├── README.md ├── data └── test-demo │ ├── T49QGF_20191017_6Bands_Urban_Subs.tif │ ├── T49QGF_20191017_6Bands_Urban_Subs_water.tif │ └── val_sam.csv ├── dataloader ├── img_aug.py ├── loader.ipynb ├── path_io.py ├── tfrecord_io.py └── tfrecord_writter.ipynb ├── environmental.yml ├── figures ├── cloudy │ ├── cloudy-awei.png │ ├── cloudy-obia.png │ ├── cloudy-scene.png │ └── cloudy-watnet.png ├── dataset.png ├── label_sam_1.png ├── label_sam_2.png ├── mountain │ ├── mountain-awei.png │ ├── mountain-obia.png │ ├── mountain-scene.png │ └── mountain-watnet.png ├── urban │ ├── urban-awei.png │ ├── urban-mndwi.png │ ├── urban-scene.png │ └── urban-watnet.png └── watnet_structure.png ├── model ├── base_model │ ├── mobilenetv2.py │ └── xception.py ├── pretrained │ └── watnet.h5 └── seg_model │ ├── deeplabv3_plus.py │ ├── deepwatermapv2.py │ └── watnet.py ├── notebooks ├── config.py ├── infer_demo.ipynb ├── metric_plot.ipynb ├── model_eval.ipynb └── trainer.ipynb ├── utils ├── acc_patch.py ├── acc_pixel.py ├── geotif_io.py ├── imgPatch.py └── imgShow.py └── watnet_infer.py /.gitignore: -------------------------------------------------------------------------------- 1 | data/dset-s2/* 2 | data/tfrecord-s2/* 3 | data/test-data 4 | result/ 5 | data/visual-data-gis 6 | model/pretrained/* 7 | !model/pretrained/watnet* 8 | .vscode 9 | .DS_Store 10 | *__pycache__/ 11 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | [![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.5205674.svg)](https://doi.org/10.5281/zenodo.5205674) 2 | 3 | # WatNet: 4 | 5 | - A deep ConvNet for surface water mapping based on Sentinel-2 image 6 | 7 | ## -- Model 8 | - We use Deeplabv3+ with MobileNetv2 backbone as the main model structure, additionly, some simple yet effective modifications are designed for improving the satellite image-based surface water mapping. 9 | 10 | ![watnet](figures/watnet_structure.png) 11 | 12 | ## -- DataSet 13 | - Surface water dataset for Deep learning could be downloaded from Zenodo [**[Link]**](https://doi.org/10.5281/zenodo.5205674). 14 | 15 | ![dataset](figures/dataset.png) 16 | |Labeling example 1:|Labeling example 2:| 17 | |:--|:--| 18 | |![example_1](figures/label_sam_1.png)|![example_2](figures/label_sam_2.png)| 19 | 20 | ## -- Performance 21 | - Examples for surface water mapping 22 | 23 | **Urban region** 24 | |Urban scene|AWEI|MNDWI|WatNet| 25 | |:--|:--|:--|:--| 26 | ||||| 27 | 28 | **Cloudy region** 29 | |Cloudy scene|AWEI|OBIA|WatNet| 30 | |:--|:--|:--|:--| 31 | ||||| 32 | 33 | **Mountainous region** 34 | |Mountain scene|AWEI|OBIA|WatNet| 35 | |:--|:--|:--|:--| 36 | ||||| 37 | 38 | 39 | ## **-- How to use the trained WatNet?** 40 | 41 | ### -- Step 1 42 | - Enter the following commands for downloading the code files, and then configure the python and deep learning environment. The deep learning software used in this repo is [Tensorflow 2.5](https://www.tensorflow.org/). 43 | 44 | ~~~console 45 | git clone https://github.com/xinluo2018/WatNet.git 46 | ~~~ 47 | 48 | ### -- Step 2 49 | - Download Sentinel-2 images, and select four visible-near infrared 10-m bands and two 20-m shortwave infrared bands, which corresponding to the band number of 2, 3, 4, 8, 11, and 12 of sentinel-2 image. 50 | 51 | ### -- Step 3 52 | - Add the prepared sentinel-2 image (6 bands) to the **_data/test-demo_** directory, modify the data name in the **_notebooks/infer_demo.ipynb_** file, then running the code file: **_notebooks/infer_demo.ipynb_** and surface water map can be generated. 53 | - Users also can specify surface water mapping by using the watnet_infer.py, specifically, 54 | - --- funtional API: 55 | ~~~python 56 | from watnet_infer import watnet_infer 57 | water_map = watnet_infer(rsimg) # full example in notebooks/infer_demo.ipynb. 58 | ~~~ 59 | - --- command line API: 60 | ~~~console 61 | python watnet_infer.py data/test-demo/*.tif -o data/test-demo/result 62 | ~~~ 63 | 64 | ## **-- How to train the WatNet?** 65 | 66 | - With the Dataset, the user can train the WatNet through running the code file **_train/trainer.ipynb_**. Since [**tfrecords**](https://www.tensorflow.org/tutorials/load_data/tfrecord?hl=zh-tw) format data is required in the model training, the user should convert the .tif format dataset to .tfrecords format dataset by running code file **_dataloader/tfrecord_writter.ipynb_** firstly. 67 | 68 | 69 | ## -- Citation 70 | 71 | - Xin Luo, Xiaohua Tong, Zhongwen Hu. An applicable and automatic method for earth surface water mapping based on multispectral images. International Journal of Applied Earth Observation and Geoinformation, 2021, 103, 102472. [[**Link**](https://www.sciencedirect.com/science/article/pii/S0303243421001793)] 72 | ``` 73 | @article{luo2021_watnet, 74 | title = {An applicable and automatic method for earth surface water mapping based on multispectral images}, 75 | author = {Xin Luo and Xiaohua Tong and Zhongwen Hu}, 76 | journal = {International Journal of Applied Earth Observation and Geoinformation}, 77 | volume = {103}, 78 | pages = {102472}, 79 | year = {2021} 80 | } 81 | ``` 82 | 83 | 84 | ## -- Acknowledgement 85 | - We thanks the authors for providing some of the code in this repo: 86 | [deeplabv3_plus](https://github.com/luyanger1799/amazing-semantic-segmentation) and [deepwatmapv2](https://github.com/isikdogan/deepwatermap) 87 | 88 | -------------------------------------------------------------------------------- /data/test-demo/T49QGF_20191017_6Bands_Urban_Subs.tif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/data/test-demo/T49QGF_20191017_6Bands_Urban_Subs.tif -------------------------------------------------------------------------------- /data/test-demo/T49QGF_20191017_6Bands_Urban_Subs_water.tif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/data/test-demo/T49QGF_20191017_6Bands_Urban_Subs_water.tif -------------------------------------------------------------------------------- /data/test-demo/val_sam.csv: -------------------------------------------------------------------------------- 1 | row,col,label 2 | 14,2207,1 3 | 14,2208,1 4 | 14,2206,1 5 | 15,2207,1 6 | 15,2208,1 7 | 15,2206,1 8 | 16,2206,1 9 | 16,2207,1 10 | 16,2208,1 11 | 316,1621,1 12 | 316,1622,1 13 | 316,1620,1 14 | 317,1620,1 15 | 317,1621,1 16 | 317,1622,1 17 | 318,1621,1 18 | 318,1622,1 19 | 318,1620,1 20 | 319,1620,1 21 | 319,1621,1 22 | 319,1622,1 23 | 826,238,1 24 | 826,237,1 25 | 827,238,1 26 | 827,237,1 27 | 828,237,1 28 | 828,238,1 29 | 990,2079,1 30 | 990,2078,1 31 | 990,2077,1 32 | 990,2080,1 33 | 991,2077,1 34 | 991,2078,1 35 | 991,2079,1 36 | 991,2080,1 37 | 996,1490,1 38 | 996,1489,1 39 | 997,1490,1 40 | 997,1489,1 41 | 998,1489,1 42 | 998,1490,1 43 | 1071,425,1 44 | 1071,424,1 45 | 1071,426,1 46 | 1072,425,1 47 | 1072,424,1 48 | 1072,426,1 49 | 1073,426,1 50 | 1073,425,1 51 | 1073,424,1 52 | 1074,426,1 53 | 1074,425,1 54 | 1074,424,1 55 | 1075,426,1 56 | 1075,425,1 57 | 1075,424,1 58 | 1076,424,1 59 | 1076,426,1 60 | 1076,425,1 61 | 1077,425,1 62 | 1077,424,1 63 | 1077,426,1 64 | 1089,836,1 65 | 1089,835,1 66 | 1090,835,1 67 | 1090,836,1 68 | 1091,835,1 69 | 1091,836,1 70 | 1116,350,1 71 | 1116,348,1 72 | 1116,349,1 73 | 1117,350,1 74 | 1117,349,1 75 | 1117,348,1 76 | 1118,349,1 77 | 1118,350,1 78 | 1118,348,1 79 | 1119,349,1 80 | 1119,348,1 81 | 1119,350,1 82 | 1153,1818,1 83 | 1153,1819,1 84 | 1153,1817,1 85 | 1154,1819,1 86 | 1154,1818,1 87 | 1154,1817,1 88 | 1155,1818,1 89 | 1155,1819,1 90 | 1155,1817,1 91 | 1169,1058,1 92 | 1169,1057,1 93 | 1170,1058,1 94 | 1170,1057,1 95 | 1171,1057,1 96 | 1171,1058,1 97 | 1212,43,1 98 | 1212,44,1 99 | 1212,45,1 100 | 1212,42,1 101 | 1213,43,1 102 | 1213,44,1 103 | 1213,45,1 104 | 1213,42,1 105 | 1286,1084,1 106 | 1286,1083,1 107 | 1287,1083,1 108 | 1287,1084,1 109 | 1311,613,1 110 | 1311,615,1 111 | 1311,614,1 112 | 1312,613,1 113 | 1312,615,1 114 | 1312,614,1 115 | 1313,615,1 116 | 1313,613,1 117 | 1313,614,1 118 | 1314,613,1 119 | 1314,615,1 120 | 1314,614,1 121 | 1356,2270,1 122 | 1356,2269,1 123 | 1356,1402,1 124 | 1356,2271,1 125 | 1357,2269,1 126 | 1357,2270,1 127 | 1357,2271,1 128 | 1357,1402,1 129 | 1358,1402,1 130 | 1358,2270,1 131 | 1358,2269,1 132 | 1358,2271,1 133 | 1405,1587,1 134 | 1405,1586,1 135 | 1406,1586,1 136 | 1406,1587,1 137 | 1407,1587,1 138 | 1407,1586,1 139 | 1408,1586,1 140 | 1408,1587,1 141 | 1422,366,1 142 | 1422,367,1 143 | 1423,367,1 144 | 1423,366,1 145 | 1424,367,1 146 | 1424,366,1 147 | 1425,366,1 148 | 1425,367,1 149 | 1430,2218,1 150 | 1430,2219,1 151 | 1430,2217,1 152 | 1431,2218,1 153 | 1431,2219,1 154 | 1431,2217,1 155 | 1432,2218,1 156 | 1432,2219,1 157 | 1432,2217,1 158 | 1433,2218,1 159 | 1433,2217,1 160 | 1433,2219,1 161 | 1489,1111,1 162 | 1489,1112,1 163 | 1489,1113,1 164 | 1490,1112,1 165 | 1490,1113,1 166 | 1490,1111,1 167 | 1491,1112,1 168 | 1491,1113,1 169 | 1491,1111,1 170 | 1586,209,1 171 | 1587,209,1 172 | 1588,209,1 173 | 1672,2005,1 174 | 1672,2004,1 175 | 1672,2003,1 176 | 1672,2007,1 177 | 1672,2006,1 178 | 1673,2005,1 179 | 1673,2006,1 180 | 1673,2007,1 181 | 1673,2003,1 182 | 1673,2004,1 183 | 1674,2007,1 184 | 1674,2005,1 185 | 1674,2004,1 186 | 1674,2003,1 187 | 1674,2006,1 188 | 1675,2004,1 189 | 1675,2005,1 190 | 1675,2006,1 191 | 1675,2007,1 192 | 1675,2003,1 193 | 1676,2004,1 194 | 1676,2003,1 195 | 1676,2006,1 196 | 1676,2007,1 197 | 1676,2005,1 198 | 1677,2005,1 199 | 1677,2007,1 200 | 1677,2004,1 201 | 1677,2003,1 202 | 1677,2006,1 203 | 1678,2005,1 204 | 1678,2004,1 205 | 1678,2003,1 206 | 1678,2006,1 207 | 1678,2007,1 208 | 1748,1012,1 209 | 1748,1011,1 210 | 1749,1012,1 211 | 1749,1011,1 212 | 1750,1012,1 213 | 1750,1011,1 214 | 1842,1062,1 215 | 1842,1063,1 216 | 1842,1061,1 217 | 1843,1061,1 218 | 1843,1062,1 219 | 1843,1063,1 220 | 1929,1436,1 221 | 1929,1435,1 222 | 1930,1436,1 223 | 1930,1435,1 224 | 1968,1822,1 225 | 1968,1823,1 226 | 1968,1824,1 227 | 1969,1823,1 228 | 1969,1824,1 229 | 1969,1822,1 230 | 1970,1822,1 231 | 1970,1823,1 232 | 1970,1824,1 233 | 1971,1823,1 234 | 1971,1824,1 235 | 1971,1822,1 236 | 35,1440,0 237 | 35,1441,0 238 | 35,1439,0 239 | 36,1440,0 240 | 36,1441,0 241 | 36,1439,0 242 | 37,1439,0 243 | 37,1440,0 244 | 37,1441,0 245 | 38,1440,0 246 | 38,1441,0 247 | 38,1439,0 248 | 39,1439,0 249 | 39,1440,0 250 | 39,1441,0 251 | 40,1440,0 252 | 40,1441,0 253 | 40,1439,0 254 | 109,1493,0 255 | 109,1494,0 256 | 109,1492,0 257 | 110,1492,0 258 | 110,1493,0 259 | 110,1494,0 260 | 111,1493,0 261 | 111,1494,0 262 | 111,1492,0 263 | 112,1492,0 264 | 112,1493,0 265 | 112,1494,0 266 | 113,1493,0 267 | 113,1494,0 268 | 113,1492,0 269 | 114,1492,0 270 | 114,1493,0 271 | 114,1494,0 272 | 138,1802,0 273 | 138,1803,0 274 | 138,1804,0 275 | 138,1801,0 276 | 139,1801,0 277 | 139,1802,0 278 | 139,1803,0 279 | 139,1804,0 280 | 140,1802,0 281 | 140,1803,0 282 | 140,1804,0 283 | 140,1801,0 284 | 141,1801,0 285 | 141,1802,0 286 | 141,1803,0 287 | 141,1804,0 288 | 278,1004,0 289 | 303,1034,0 290 | 377,1257,0 291 | 482,1351,0 292 | 483,1318,0 293 | 543,513,0 294 | 543,514,0 295 | 543,515,0 296 | 543,516,0 297 | 544,514,0 298 | 544,515,0 299 | 544,516,0 300 | 544,513,0 301 | 545,513,0 302 | 545,514,0 303 | 545,515,0 304 | 545,516,0 305 | 546,514,0 306 | 546,515,0 307 | 546,516,0 308 | 546,513,0 309 | 547,513,0 310 | 547,514,0 311 | 547,515,0 312 | 547,516,0 313 | 697,2218,0 314 | 697,2219,0 315 | 697,2217,0 316 | 698,2217,0 317 | 698,2218,0 318 | 698,2219,0 319 | 699,2218,0 320 | 699,2219,0 321 | 699,2217,0 322 | 700,2217,0 323 | 700,2218,0 324 | 700,2219,0 325 | 701,2218,0 326 | 701,2219,0 327 | 701,2217,0 328 | 822,324,0 329 | 822,325,0 330 | 822,323,0 331 | 822,326,0 332 | 823,325,0 333 | 823,326,0 334 | 823,324,0 335 | 823,323,0 336 | 824,324,0 337 | 824,325,0 338 | 824,323,0 339 | 824,326,0 340 | 825,323,0 341 | 825,325,0 342 | 825,326,0 343 | 825,324,0 344 | 826,325,0 345 | 826,323,0 346 | 826,326,0 347 | 826,324,0 348 | 827,325,0 349 | 827,326,0 350 | 827,324,0 351 | 827,323,0 352 | 828,325,0 353 | 828,324,0 354 | 828,323,0 355 | 828,326,0 356 | 964,136,0 357 | 964,137,0 358 | 964,135,0 359 | 965,136,0 360 | 965,137,0 361 | 965,135,0 362 | 966,136,0 363 | 966,135,0 364 | 966,137,0 365 | 967,136,0 366 | 967,137,0 367 | 967,135,0 368 | 968,136,0 369 | 968,135,0 370 | 968,137,0 371 | 1094,136,0 372 | 1194,1276,0 373 | 1207,953,0 374 | 1315,415,0 375 | 1395,425,0 376 | 1659,550,0 377 | 1659,551,0 378 | 1659,549,0 379 | 1660,549,0 380 | 1660,550,0 381 | 1660,551,0 382 | 1661,550,0 383 | 1661,551,0 384 | 1661,549,0 385 | 1708,2065,0 386 | 1708,2064,0 387 | 1708,2066,0 388 | 1708,2067,0 389 | 1708,2068,0 390 | 1708,2069,0 391 | 1709,2065,0 392 | 1709,2066,0 393 | 1709,2067,0 394 | 1709,2069,0 395 | 1709,2068,0 396 | 1709,2064,0 397 | 1710,2065,0 398 | 1710,2066,0 399 | 1710,2067,0 400 | 1710,2068,0 401 | 1710,2069,0 402 | 1710,2064,0 403 | 1711,2064,0 404 | 1711,2065,0 405 | 1711,2066,0 406 | 1711,2067,0 407 | 1711,2068,0 408 | 1711,2069,0 409 | 1712,2064,0 410 | 1712,2065,0 411 | 1712,2066,0 412 | 1712,2067,0 413 | 1712,2068,0 414 | 1712,2069,0 415 | 1713,2065,0 416 | 1713,2066,0 417 | 1713,2068,0 418 | 1713,2067,0 419 | 1713,2064,0 420 | 1713,2069,0 421 | 1714,2067,0 422 | 1714,2065,0 423 | 1714,2064,0 424 | 1714,2068,0 425 | 1714,2069,0 426 | 1714,2066,0 427 | 1715,2067,0 428 | 1715,2064,0 429 | 1715,2068,0 430 | 1715,2065,0 431 | 1715,2069,0 432 | 1715,2066,0 433 | 1771,1552,0 434 | 1771,1550,0 435 | 1771,1553,0 436 | 1771,1551,0 437 | 1772,1550,0 438 | 1772,1551,0 439 | 1772,1552,0 440 | 1772,1553,0 441 | 1773,1551,0 442 | 1773,1552,0 443 | 1773,1553,0 444 | 1773,1550,0 445 | 1774,1550,0 446 | 1774,1551,0 447 | 1774,1552,0 448 | 1774,1553,0 449 | 1775,1553,0 450 | 1775,1551,0 451 | 1775,1552,0 452 | 1775,1550,0 453 | 1776,1552,0 454 | 1776,1551,0 455 | 1776,1550,0 456 | 1776,1553,0 457 | 1777,1551,0 458 | 1777,1552,0 459 | 1777,1553,0 460 | 1777,1550,0 461 | 1796,304,0 462 | 1796,303,0 463 | 1797,304,0 464 | 1797,303,0 465 | 1798,303,0 466 | 1798,304,0 467 | 1851,961,0 468 | 1851,960,0 469 | 1851,959,0 470 | 1851,962,0 471 | 1852,959,0 472 | 1852,960,0 473 | 1852,961,0 474 | 1852,962,0 475 | 1853,960,0 476 | 1853,962,0 477 | 1853,961,0 478 | 1853,959,0 479 | 1854,959,0 480 | 1854,960,0 481 | 1854,961,0 482 | 1854,962,0 483 | 1859,605,0 484 | 1859,606,0 485 | 1859,607,0 486 | 1860,606,0 487 | 1860,605,0 488 | 1860,607,0 489 | 1861,606,0 490 | 1861,607,0 491 | 1861,605,0 492 | 1862,606,0 493 | 1862,607,0 494 | 1862,605,0 495 | -------------------------------------------------------------------------------- /dataloader/img_aug.py: -------------------------------------------------------------------------------- 1 | import random 2 | import tensorflow as tf 3 | 4 | def img_aug(image, truth, flip = True, rot = True, noisy = True): 5 | '''Data augmentation: noisy, filp, rotate. ''' 6 | # image = tf.convert_to_tensor(image, dtype=tf.float32) 7 | # truth = tf.convert_to_tensor(truth, dtype=tf.float32) 8 | if len(truth.shape) == 2: 9 | truth = tf.expand_dims(truth, axis=-1) 10 | if flip == True: 11 | if tf.random.uniform(()) > 0.5: 12 | if random.randint(1,2) == 1: ## horizontal or vertical mirroring 13 | image = tf.image.flip_left_right(image) 14 | truth = tf.image.flip_left_right(truth) 15 | else: 16 | image = tf.image.flip_up_down(image) 17 | truth = tf.image.flip_up_down(truth) 18 | if rot == True: 19 | if tf.random.uniform(()) > 0.5: 20 | degree = random.randint(1,3) 21 | image = tf.image.rot90(image, k=degree) 22 | truth = tf.image.rot90(truth, k=degree) 23 | if noisy == True: 24 | if tf.random.uniform(()) > 0.5: 25 | std = random.uniform(0.001, 0.05) 26 | gnoise = tf.random.normal(shape=tf.shape(image), mean=0.0, stddev=std, dtype=tf.float32) 27 | image = tf.add(image, gnoise) 28 | return image, truth 29 | 30 | 31 | -------------------------------------------------------------------------------- /dataloader/path_io.py: -------------------------------------------------------------------------------- 1 | import random 2 | import numpy as np 3 | from utils.geotif_io import readTiff 4 | 5 | def read_scene_pair(paths_scene, paths_truth): 6 | '''read data from path and 0-1 normalization 7 | ''' 8 | scenes = [] 9 | truths = [] 10 | paths_scene_pair = zip(paths_scene, paths_truth) 11 | for path_scene, path_truth in paths_scene_pair: 12 | scene,_ = readTiff(path_scene) 13 | truth,_ = readTiff(path_truth) 14 | scene = np.clip(scene/10000,0,1) # normalization 15 | scenes.append(scene) 16 | truths.append(truth) 17 | return scenes, truths 18 | 19 | def crop_patch(img, truth, width=512, height=512, _random=True): 20 | '''crop image to patch''' 21 | if _random: 22 | row_start = random.randint(0, img.shape[0]-height) 23 | col_start = random.randint(0, img.shape[1]-width) 24 | else: 25 | row_start = 0 26 | col_start = 0 27 | patch = img[row_start:row_start+height, col_start:col_start+width] 28 | truth = truth[row_start:row_start+height, col_start:col_start+width] 29 | patch, truth = patch.astype(np.float32), truth.astype(np.float32) 30 | return patch, truth 31 | 32 | def crop_patches(imgs, truths, width=512, height=512, _random=True): 33 | '''crop images to patchs 34 | augs: 35 | imgs: a list contains images 36 | truths: a list contains image truths 37 | return: 38 | cropted patches list truth list 39 | ''' 40 | patches, ptruths = [],[] 41 | for i in range(len(imgs)): 42 | patch, truth = crop_patch(img=imgs[i], truth=truths[i], \ 43 | width=width, height=height, _random=_random) 44 | patches.append(patch) 45 | ptruths.append(truth) 46 | return patches, ptruths 47 | 48 | -------------------------------------------------------------------------------- /dataloader/tfrecord_io.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | 3 | feaBand = ['blue','green','red','nir','swir1','swir2'] 4 | truBand = ['truth'] 5 | mergeBand = feaBand+truBand 6 | 7 | 8 | '''-------------write out-------------''' 9 | def int64_feature(value): 10 | """Returns an int64_list from a bool / enum / int / uint.""" 11 | return tf.train.Feature(int64_list=tf.train.Int64List(value=value)) 12 | 13 | def float_feature(value): 14 | """Returns a float_list from a float / double.""" 15 | return tf.train.Feature(float_list=tf.train.FloatList(value=value)) 16 | 17 | # Create a dictionary with features that may be relevant. 18 | def image_example(image,truth): 19 | feature = { 20 | 'bandShape': int64_feature(image[:,:,0].shape), 21 | 'blue': float_feature(image[:,:,0].flatten()), 22 | 'green': float_feature(image[:,:,1].flatten()), 23 | 'red': float_feature(image[:,:,2].flatten()), 24 | 'nir': float_feature(image[:,:,3].flatten()), 25 | 'swir1': float_feature(image[:,:,4].flatten()), 26 | 'swir2': float_feature(image[:,:,5].flatten()), 27 | 'truth': float_feature(truth.flatten()), 28 | } 29 | return tf.train.Example(features=tf.train.Features(feature=feature)) 30 | 31 | '''-------------read in-------------''' 32 | featuresDict = { 33 | 'bandShape': tf.io.FixedLenFeature([2,],dtype=tf.int64), 34 | 'blue': tf.io.VarLenFeature(dtype=tf.float32), 35 | 'green': tf.io.VarLenFeature(dtype=tf.float32), 36 | 'red': tf.io.VarLenFeature(dtype=tf.float32), 37 | 'nir': tf.io.VarLenFeature(dtype=tf.float32), 38 | 'swir1': tf.io.VarLenFeature(dtype=tf.float32), 39 | 'swir2': tf.io.VarLenFeature(dtype=tf.float32), 40 | 'truth': tf.io.VarLenFeature(dtype=tf.float32), 41 | } 42 | 43 | def parse_image(example_proto): 44 | # Parse the input tf.train.Example proto using the dictionary above. 45 | return tf.io.parse_single_example(example_proto, featuresDict) 46 | 47 | def parse_shape(example_parsed): 48 | for fea in mergeBand: 49 | example_parsed[fea] = tf.sparse.to_dense(example_parsed[fea]) 50 | example_parsed[fea] = tf.reshape(example_parsed[fea], example_parsed['bandShape']) 51 | return example_parsed 52 | 53 | def toPatchPair(inputs): 54 | inputsList = [inputs.get(key) for key in mergeBand] 55 | stacked = tf.stack(inputsList, axis=2) 56 | cropped_stacked = tf.image.random_crop( 57 | stacked, size=[512, 512, len(mergeBand)]) 58 | return cropped_stacked[:,:,:len(feaBand)], cropped_stacked[:,:,len(feaBand):] 59 | 60 | -------------------------------------------------------------------------------- /dataloader/tfrecord_writter.ipynb: -------------------------------------------------------------------------------- 1 | {"nbformat":4,"nbformat_minor":2,"metadata":{"colab":{"name":"tfrecord_writer.ipynb","provenance":[],"collapsed_sections":[]},"kernelspec":{"name":"python3","display_name":"Python 3.6.13 64-bit ('venv-tf': conda)"},"language_info":{"name":"python","version":"3.6.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"interpreter":{"hash":"540331ab3874cc423b0568d77e5ba6321eb807f34352bd2dfeeaef70fdba3ae8"}},"cells":[{"cell_type":"code","execution_count":2,"source":["# ### mount on google drive if you use colab\n","# from google.colab import drive\n","# drive.mount('/content/drive/')\n","# import os\n","# os.chdir(\"/content/drive/My Drive/WatNet/notebooks\")\n"],"outputs":[{"output_type":"stream","name":"stdout","text":["Drive already mounted at /content/drive/; to attempt to forcibly remount, call drive.mount(\"/content/drive/\", force_remount=True).\n"]}],"metadata":{"id":"g1ppHTj0WylH","colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"status":"ok","timestamp":1622636666681,"user_tz":-480,"elapsed":447,"user":{"displayName":"Xin Luo","photoUrl":"","userId":"06301970496892076570"}},"outputId":"39d90e1a-8e25-41af-e7d0-4725abb7d679"}},{"cell_type":"code","execution_count":1,"source":["import os\n","os.chdir('..')\n","import glob\n","import numpy as np\n","import tensorflow as tf\n","from utils.geotif_io import readTiff\n","from dataloader.path_io import crop_patch\n","from dataloader.tfrecord_io import image_example\n"],"outputs":[],"metadata":{"id":"gFkuYIjIYwj6","executionInfo":{"status":"ok","timestamp":1622636673866,"user_tz":-480,"elapsed":2810,"user":{"displayName":"Xin Luo","photoUrl":"","userId":"06301970496892076570"}}}},{"cell_type":"code","execution_count":3,"source":["path_tra_tfrecord_scene = 'data/tfrecord-s2/tra_scene.tfrecords'\n","path_val_tfrecord_patch = 'data/tfrecord-s2/val_patch.tfrecords'\n","path_val_tfrecord_scene = 'data/tfrecord-s2/val_scene.tfrecords'\n","tra_scene_paths = sorted(glob.glob('data/dset-s2/tra_scene/*.tif'))\n","tra_truth_paths = sorted(glob.glob('data/dset-s2/tra_truth/*.tif'))\n","tra_pair_data = list(zip(tra_scene_paths, tra_truth_paths))\n","print('tra_data length:', len(tra_pair_data))\n","val_scene_paths = sorted(glob.glob('data/dset-s2/val_scene/*.tif'))\n","val_truth_paths = sorted(glob.glob('data/dset-s2/val_truth/*.tif'))\n","val_pair_data = list(zip(val_scene_paths, val_truth_paths))\n","print('val_data length:', len(val_pair_data))\n","\n"],"outputs":[{"output_type":"stream","name":"stdout","text":["tra_data length: 64\n","val_data length: 31\n"]}],"metadata":{"id":"uhy8whHOWzps","colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"status":"ok","timestamp":1622636679077,"user_tz":-480,"elapsed":943,"user":{"displayName":"Xin Luo","photoUrl":"","userId":"06301970496892076570"}},"outputId":"d9de8cf7-773a-40ff-8dca-93dc6d354904"}},{"cell_type":"code","execution_count":4,"source":["# Trainging data: Write to a `.tfrecords` file.\n","with tf.io.TFRecordWriter(path_tra_tfrecord_scene) as writer:\n"," for path_scene, path_truth in tra_pair_data:\n"," print(path_scene)\n"," scene,_ = readTiff(path_scene)\n"," truth,_ = readTiff(path_truth)\n"," scene = np.clip(scene/10000,0,1) \n"," tf_example = image_example(scene, truth)\n"," # writer.write(tf_example.SerializeToString())\n"],"outputs":[{"output_type":"stream","name":"stdout","text":["data/dset-s2/tra_scene/S2A_L2A_20190125_N0211_R034_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190125_N0211_R034_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190125_N0211_R034_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190206_N0211_R067_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190206_N0211_R067_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190206_N0211_R067_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190314_N0211_R008_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190314_N0211_R008_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190314_N0211_R008_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190716_N0213_R063_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190716_N0213_R063_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190716_N0213_R063_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190811_N0213_R013_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190811_N0213_R013_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190811_N0213_R013_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190817_N0213_R089_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190817_N0213_R089_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190817_N0213_R089_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190817_N0213_R089_6Bands_S4.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190830_N0213_R133_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190830_N0213_R133_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2A_L2A_20190830_N0213_R133_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20181226_N0211_R102_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20181226_N0211_R102_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20181226_N0211_R102_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190225_N0211_R117_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190225_N0211_R117_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190225_N0211_R117_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190430_N0211_R035_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190430_N0211_R035_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190430_N0211_R035_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190506_N0212_R126_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190506_N0212_R126_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190506_N0212_R126_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190607_N0212_R012_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190607_N0212_R012_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190607_N0212_R012_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190620_N0212_R053_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190620_N0212_R053_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190620_N0212_R053_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190801_N0213_R084_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190801_N0213_R084_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190801_N0213_R084_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190807_N0213_R018_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190807_N0213_R018_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190807_N0213_R018_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190811_N0213_R079_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190811_N0213_R079_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190811_N0213_R079_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190818_N0213_R035_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190818_N0213_R035_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190818_N0213_R035_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190831_N0213_R082_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190831_N0213_R082_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190831_N0213_R082_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190904_N0213_R132_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190904_N0213_R132_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190904_N0213_R132_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190912_N0213_R102_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190912_N0213_R102_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20190912_N0213_R102_6Bands_S3.tif\n","data/dset-s2/tra_scene/S2B_L2A_20191023_N0213_R121_6Bands_S1.tif\n","data/dset-s2/tra_scene/S2B_L2A_20191023_N0213_R121_6Bands_S2.tif\n","data/dset-s2/tra_scene/S2B_L2A_20191023_N0213_R121_6Bands_S3.tif\n"]}],"metadata":{"id":"4vu-8itudUwx"}},{"cell_type":"code","execution_count":5,"source":["# Val data: Write to a `_scene.tfrecords` file.\n","with tf.io.TFRecordWriter(path_val_tfrecord_scene) as writer:\n"," for path_scene, path_truth in val_pair_data:\n"," print(path_scene)\n"," scene,_ = readTiff(path_scene)\n"," truth,_ = readTiff(path_truth)\n"," scene = np.clip(scene/10000,0,1)\n"," tf_example = image_example(scene, truth)\n"," # writer.write(tf_example.SerializeToString())\n"],"outputs":[{"output_type":"stream","name":"stdout","text":["data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n"]}],"metadata":{}},{"cell_type":"code","execution_count":6,"source":["## Validation data: Write to a `_patch.tfrecords` file.\n","with tf.io.TFRecordWriter(path_val_tfrecord_patch) as writer:\n"," for i in range(10): # random croping 10 times for each scene\n"," for path_scene, path_truth in val_pair_data:\n"," print(path_scene)\n"," scene,_ = readTiff(path_scene)\n"," truth,_ = readTiff(path_truth)\n"," scene = np.clip(scene/10000,0,1)\n"," patch, truth = crop_patch(img=scene, truth=truth)\n"," tf_example = image_example(patch, truth)\n"," # writer.write(tf_example.SerializeToString())\n"],"outputs":[{"output_type":"stream","name":"stdout","text":["data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190318_N0211_R061_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190426_N0211_R053_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190429_N0211_R096_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190508_N0212_R078_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190518_N0212_R073_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190724_N0213_R033_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S1.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S2.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S3.tif\n","data/dset-s2/val_scene/S2A_L2A_20190725_N0213_R054_6Bands_S4.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190303_N0211_R065_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20190620_N0212_R047_6Bands_S3.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S1.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S2.tif\n","data/dset-s2/val_scene/S2B_L2A_20191015_N0213_R007_6Bands_S3.tif\n"]}],"metadata":{"id":"A4_vq76_rb_O","executionInfo":{"status":"ok","timestamp":1622636805400,"user_tz":-480,"elapsed":368,"user":{"displayName":"Xin Luo","photoUrl":"","userId":"06301970496892076570"}},"tags":[]}}]} -------------------------------------------------------------------------------- /environmental.yml: -------------------------------------------------------------------------------- 1 | name: venv-tf 2 | channels: 3 | - conda-forge 4 | - defaults 5 | dependencies: 6 | - gdal=3.2.1=py36h5adf297_0 7 | - geos=3.8.1=he1b5a44_0 8 | - geotiff=1.6.0=h5d11630_3 9 | - gettext=0.19.8.1=h0b5b191_1005 10 | - matplotlib=3.2.2=1 11 | - numpy=1.19.5=py36h2aa4a07_1 12 | - pandas=1.1.5=py36h284efc9_0 13 | - proj=7.1.1=h966b41f_3 14 | - python=3.6.13=hffdb5ce_0_cpython 15 | - scikit-learn=0.24.2=py36h2fdd933_0 16 | - scipy=1.5.3=py36h9e8f40b_0 17 | - six=1.16.0=pyh6c4a22f_0 18 | - pip: 19 | - tensorflow==2.5.0 20 | prefix: /home/yons/miniconda3/envs/venv-tf 21 | -------------------------------------------------------------------------------- /figures/cloudy/cloudy-awei.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/cloudy/cloudy-awei.png -------------------------------------------------------------------------------- /figures/cloudy/cloudy-obia.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/cloudy/cloudy-obia.png -------------------------------------------------------------------------------- /figures/cloudy/cloudy-scene.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/cloudy/cloudy-scene.png -------------------------------------------------------------------------------- /figures/cloudy/cloudy-watnet.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/cloudy/cloudy-watnet.png -------------------------------------------------------------------------------- /figures/dataset.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/dataset.png -------------------------------------------------------------------------------- /figures/label_sam_1.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/label_sam_1.png -------------------------------------------------------------------------------- /figures/label_sam_2.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/label_sam_2.png -------------------------------------------------------------------------------- /figures/mountain/mountain-awei.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/mountain/mountain-awei.png -------------------------------------------------------------------------------- /figures/mountain/mountain-obia.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/mountain/mountain-obia.png -------------------------------------------------------------------------------- /figures/mountain/mountain-scene.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/mountain/mountain-scene.png -------------------------------------------------------------------------------- /figures/mountain/mountain-watnet.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/mountain/mountain-watnet.png -------------------------------------------------------------------------------- /figures/urban/urban-awei.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/urban/urban-awei.png -------------------------------------------------------------------------------- /figures/urban/urban-mndwi.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/urban/urban-mndwi.png -------------------------------------------------------------------------------- /figures/urban/urban-scene.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/urban/urban-scene.png -------------------------------------------------------------------------------- /figures/urban/urban-watnet.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/urban/urban-watnet.png -------------------------------------------------------------------------------- /figures/watnet_structure.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/figures/watnet_structure.png -------------------------------------------------------------------------------- /model/base_model/mobilenetv2.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | 3 | ##### MobileNetV2 4 | relu6 = tf.keras.layers.ReLU(6.) 5 | def _conv_block(inputs, filters, kernel, strides): 6 | x = tf.keras.layers.Conv2D(filters, kernel, padding='same', strides=strides)(inputs) 7 | x = tf.keras.layers.BatchNormalization()(x) 8 | return relu6(x) 9 | def _bottleneck(inputs, filters, kernel, t, s, r=False): 10 | tchannel = inputs.shape[-1] * t 11 | x = _conv_block(inputs, tchannel, (1, 1), (1, 1)) 12 | x = tf.keras.layers.DepthwiseConv2D(kernel, strides=(s, s), depth_multiplier=1, padding='same')(x) 13 | x = tf.keras.layers.BatchNormalization()(x) 14 | x = relu6(x) 15 | x = tf.keras.layers.Conv2D(filters, (1, 1), strides=(1, 1), padding='same')(x) # 降维，改层为瓶颈层 16 | x = tf.keras.layers.BatchNormalization()(x) 17 | if r: 18 | x = tf.keras.layers.add([x, inputs]) 19 | return x 20 | 21 | def _inverted_residual_block(inputs, filters, kernel, t, strides, n): 22 | x = _bottleneck(inputs, filters, kernel, t, strides) 23 | for i in range(1, n): 24 | x = _bottleneck(x, filters, kernel, t, 1, True) 25 | return x 26 | 27 | def MobileNetV2(input_shape, nclasses=2): 28 | 29 | """ 30 | # Arguments 31 | input_shape: An integer or tuple/list of 3 integers, shape 32 | of input tensor. 33 | classes: Integer, number of classes. 34 | # Returns 35 | MobileNetv2 model. 36 | """ 37 | 38 | inputs = tf.keras.layers.Input(shape=input_shape, name='input') 39 | x = _conv_block(inputs, 32, (3, 3), strides=(2, 2)) # 0.5*size n_layers = 1 40 | 41 | x = _inverted_residual_block(x, 16, (3, 3), t=1, strides=1, n=1) # n_layers = 3 42 | x = _inverted_residual_block(x, 24, (3, 3), t=6, strides=2, n=2) # 0.5*size, n_layers = 6 43 | 44 | x = _inverted_residual_block(x, 32, (3, 3), t=6, strides=2, n=3) # 0.5*size, n_layers = 9 45 | x = _inverted_residual_block(x, 64, (3, 3), t=6, strides=2, n=4) # 0.5*size, n_layers = 12 46 | x = _inverted_residual_block(x, 96, (3, 3), t=6, strides=1, n=3) # n_layers = 9 47 | x = _inverted_residual_block(x, 160, (3, 3), t=6, strides=2, n=3) # 0.5*size, n_layers = 9 48 | x = _inverted_residual_block(x, 320, (3, 3), t=6, strides=1, n=1) # n_layers = 3 49 | 50 | x = _conv_block(x, 1280, (1, 1), strides=(1, 1)) # n_layers = 1 51 | x = tf.keras.layers.GlobalAveragePooling2D()(x) 52 | x = tf.keras.layers.Reshape((1, 1, 1280))(x) 53 | x = tf.keras.layers.Dropout(0.3, name='Dropout')(x) 54 | x = tf.keras.layers.Conv2D(nclasses, (1, 1), padding='same')(x) # n_layers = 1 55 | x = tf.keras.layers.Activation('softmax', name='final_activation')(x) 56 | output = tf.keras.layers.Reshape((nclasses,), name='output')(x) 57 | model = tf.keras.models.Model(inputs, output) 58 | return model 59 | 60 | # model = MobileNetV2(input_shape=(512,512,6), nclasses=2) 61 | # model.summary() 62 | 63 | -------------------------------------------------------------------------------- /model/base_model/xception.py: -------------------------------------------------------------------------------- 1 | 2 | '''reference: https://github.com/luyanger1799/Amazing-Semantic-Segmentation 3 | ''' 4 | 5 | import os 6 | import sys 7 | sys.path.append(os.getcwd()) 8 | import tensorflow as tf 9 | import tensorflow.keras.layers as layers 10 | import tensorflow.keras.backend as backend 11 | 12 | class Xception(): 13 | def __init__(self, version='Xception', dilation=None, **kwargs): 14 | """ 15 | The Xception is used to structure DeepLabV3Plus based on Tensorflow. 16 | :param dilation: Whether to use dilation strategy 17 | """ 18 | super(Xception, self).__init__(**kwargs) 19 | self.version = version 20 | if dilation is None: 21 | self.strides = [2, 2] 22 | else: 23 | self.strides = [2 if dilation[0] == 1 else 1] + [2 if dilation[1] == 1 else 1] 24 | assert len(self.strides) == 2 25 | assert version in ['Xception', 'Xception-DeepLab'] 26 | 27 | def __call__(self, inputs, output_stages='c5', **kwargs): 28 | """ 29 | call for Xception or Xception-DeepLab. 30 | :param inputs: a 4-D tensor. 31 | :param output_stages: str or a list of str containing the output stages. 32 | :param kwargs: other parameters. 33 | :return: the output of different stages. 34 | """ 35 | strides = self.strides 36 | if self.version == 'Xception-DeepLab': 37 | rm_pool = True 38 | num_middle_flow = 16 39 | else: 40 | rm_pool = False 41 | num_middle_flow = 8 42 | 43 | channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1 44 | 45 | x = layers.Conv2D(32, (3, 3), 46 | strides=(2, 2), 47 | use_bias=False, 48 | padding='same', 49 | name='block1_conv1')(inputs) 50 | x = layers.BatchNormalization(axis=channel_axis, name='block1_conv1_bn')(x) 51 | x = layers.Activation('relu', name='block1_conv1_act')(x) 52 | x = layers.Conv2D(64, (3, 3), use_bias=False, padding='same', name='block1_conv2')(x) 53 | x = layers.BatchNormalization(axis=channel_axis, name='block1_conv2_bn')(x) 54 | x = layers.Activation('relu', name='block1_conv2_act')(x) 55 | 56 | residual = layers.Conv2D(128, (1, 1), 57 | strides=(2, 2), 58 | padding='same', 59 | use_bias=False)(x) 60 | residual = layers.BatchNormalization(axis=channel_axis)(residual) 61 | 62 | x = layers.SeparableConv2D(128, (3, 3), 63 | padding='same', 64 | use_bias=False, 65 | name='block2_sepconv1')(x) 66 | x = layers.BatchNormalization(axis=channel_axis, name='block2_sepconv1_bn')(x) 67 | x = layers.Activation('relu', name='block2_sepconv2_act')(x) 68 | x = layers.SeparableConv2D(128, (3, 3), 69 | padding='same', 70 | use_bias=False, 71 | name='block2_sepconv2')(x) 72 | x = layers.BatchNormalization(axis=channel_axis, name='block2_sepconv2_bn')(x) 73 | 74 | x = layers.MaxPooling2D((3, 3), 75 | strides=(2, 2), 76 | padding='same', 77 | name='block2_pool')(x) 78 | x = layers.add([x, residual]) 79 | c1 = x 80 | 81 | residual = layers.Conv2D(256, (1, 1), strides=(2, 2), 82 | padding='same', use_bias=False)(x) 83 | residual = layers.BatchNormalization(axis=channel_axis)(residual) 84 | 85 | x = layers.Activation('relu', name='block3_sepconv1_act')(x) 86 | x = layers.SeparableConv2D(256, (3, 3), 87 | padding='same', 88 | use_bias=False, 89 | name='block3_sepconv1')(x) 90 | x = layers.BatchNormalization(axis=channel_axis, name='block3_sepconv1_bn')(x) 91 | x = layers.Activation('relu', name='block3_sepconv2_act')(x) 92 | x = layers.SeparableConv2D(256, (3, 3), 93 | padding='same', 94 | use_bias=False, 95 | name='block3_sepconv2')(x) 96 | x = layers.BatchNormalization(axis=channel_axis, name='block3_sepconv2_bn')(x) 97 | 98 | if rm_pool: 99 | x = layers.Activation('relu', name='block3_sepconv3_act')(x) 100 | x = layers.SeparableConv2D(256, (3, 3), 101 | strides=(2, 2), 102 | padding='same', 103 | use_bias=False, 104 | name='block3_sepconv3')(x) 105 | x = layers.BatchNormalization(axis=channel_axis, name='block3_sepconv3_bn')(x) 106 | else: 107 | x = layers.MaxPooling2D((3, 3), strides=(2, 2), 108 | padding='same', 109 | name='block3_pool')(x) 110 | x = layers.add([x, residual]) 111 | c2 = x 112 | 113 | residual = layers.Conv2D(728, (1, 1), 114 | strides=strides[0], 115 | padding='same', 116 | use_bias=False)(x) 117 | residual = layers.BatchNormalization(axis=channel_axis)(residual) 118 | 119 | x = layers.Activation('relu', name='block4_sepconv1_act')(x) 120 | x = layers.SeparableConv2D(728, (3, 3), 121 | padding='same', 122 | use_bias=False, 123 | name='block4_sepconv1')(x) 124 | x = layers.BatchNormalization(axis=channel_axis, name='block4_sepconv1_bn')(x) 125 | x = layers.Activation('relu', name='block4_sepconv2_act')(x) 126 | x = layers.SeparableConv2D(728, (3, 3), 127 | padding='same', 128 | use_bias=False, 129 | name='block4_sepconv2')(x) 130 | x = layers.BatchNormalization(axis=channel_axis, name='block4_sepconv2_bn')(x) 131 | 132 | if rm_pool: 133 | x = layers.Activation('relu', name='block4_sepconv3_act')(x) 134 | x = layers.SeparableConv2D(728, (3, 3), 135 | strides=strides[0], 136 | padding='same', 137 | use_bias=False, 138 | name='block4_sepconv3')(x) 139 | x = layers.BatchNormalization(axis=channel_axis, name='block4_sepconv3_bn')(x) 140 | else: 141 | x = layers.MaxPooling2D((3, 3), strides=(2, 2), 142 | padding='same', 143 | name='block4_pool')(x) 144 | x = layers.add([x, residual]) 145 | c3 = x 146 | 147 | for i in range(num_middle_flow): 148 | residual = x 149 | prefix = 'block' + str(i + 5) 150 | 151 | x = layers.Activation('relu', name=prefix + '_sepconv1_act')(x) 152 | x = layers.SeparableConv2D(728, (3, 3), 153 | padding='same', 154 | use_bias=False, 155 | name=prefix + '_sepconv1')(x) 156 | x = layers.BatchNormalization(axis=channel_axis, 157 | name=prefix + '_sepconv1_bn')(x) 158 | x = layers.Activation('relu', name=prefix + '_sepconv2_act')(x) 159 | x = layers.SeparableConv2D(728, (3, 3), 160 | padding='same', 161 | use_bias=False, 162 | name=prefix + '_sepconv2')(x) 163 | x = layers.BatchNormalization(axis=channel_axis, 164 | name=prefix + '_sepconv2_bn')(x) 165 | x = layers.Activation('relu', name=prefix + '_sepconv3_act')(x) 166 | x = layers.SeparableConv2D(728, (3, 3), 167 | padding='same', 168 | use_bias=False, 169 | name=prefix + '_sepconv3')(x) 170 | x = layers.BatchNormalization(axis=channel_axis, 171 | name=prefix + '_sepconv3_bn')(x) 172 | 173 | x = layers.add([x, residual]) 174 | c4 = x 175 | 176 | residual = layers.Conv2D(1024, (1, 1), strides=strides[1], 177 | padding='same', use_bias=False)(x) 178 | residual = layers.BatchNormalization(axis=channel_axis)(residual) 179 | 180 | id = 5 + num_middle_flow 181 | x = layers.Activation('relu', name='block{id}_sepconv1_act'.format(id=id))(x) 182 | x = layers.SeparableConv2D(728, (3, 3), 183 | padding='same', 184 | use_bias=False, 185 | name='block{id}_sepconv1'.format(id=id))(x) 186 | x = layers.BatchNormalization(axis=channel_axis, name='block{id}_sepconv1_bn'.format(id=id))(x) 187 | x = layers.Activation('relu', name='block{id}_sepconv2_act'.format(id=id))(x) 188 | x = layers.SeparableConv2D(1024, (3, 3), 189 | padding='same', 190 | use_bias=False, 191 | name='block{id}_sepconv2'.format(id=id))(x) 192 | x = layers.BatchNormalization(axis=channel_axis, name='block{id}_sepconv2_bn'.format(id=id))(x) 193 | 194 | if rm_pool: 195 | x = layers.Activation('relu', name='block{id}_sepconv3_act'.format(id=id))(x) 196 | x = layers.SeparableConv2D(1024, (3, 3), 197 | strides=strides[1], 198 | padding='same', 199 | use_bias=False, 200 | name='block{id}_sepconv3'.format(id=id))(x) 201 | x = layers.BatchNormalization(axis=channel_axis, name='block{id}_sepconv3_bn'.format(id=id))(x) 202 | else: 203 | x = layers.MaxPooling2D((3, 3), 204 | strides=(2, 2), 205 | padding='same', 206 | name='block{id}_pool'.format(id=id))(x) 207 | x = layers.add([x, residual]) 208 | 209 | x = layers.SeparableConv2D(1536, (3, 3), 210 | padding='same', 211 | use_bias=False, 212 | name='block{id}_sepconv1'.format(id=id + 1))(x) 213 | x = layers.BatchNormalization(axis=channel_axis, name='block{id}_sepconv1_bn'.format(id=id + 1))(x) 214 | x = layers.Activation('relu', name='block{id}_sepconv1_act'.format(id=id + 1))(x) 215 | 216 | if self.version == 'Xception-DeepLab': 217 | x = layers.SeparableConv2D(1536, (3, 3), 218 | padding='same', 219 | use_bias=False, 220 | name='block{id}_sepconv1_1'.format(id=id + 1))(x) 221 | x = layers.BatchNormalization(axis=channel_axis, name='block{id}_sepconv1_1_bn'.format(id=id + 1))(x) 222 | x = layers.Activation('relu', name='block{id}_sepconv1_1_act'.format(id=id + 1))(x) 223 | 224 | x = layers.SeparableConv2D(2048, (3, 3), 225 | padding='same', 226 | use_bias=False, 227 | name='block{id}_sepconv2'.format(id=id + 1))(x) 228 | x = layers.BatchNormalization(axis=channel_axis, name='block{id}_sepconv2_bn'.format(id=id + 1))(x) 229 | x = layers.Activation('relu', name='block{id}_sepconv2_act'.format(id=id + 1))(x) 230 | 231 | c5 = x 232 | 233 | self.outputs = {'c1': c1, 234 | 'c2': c2, 235 | 'c3': c3, 236 | 'c4': c4, 237 | 'c5': c5} 238 | 239 | if type(output_stages) is not list: 240 | return self.outputs[output_stages] 241 | else: 242 | return [self.outputs[ci] for ci in output_stages] 243 | 244 | 245 | # input = tf.ones([4, 256, 256, 4],tf.float32) 246 | # model = Xception() 247 | # oupt = model(inputs=input, output_stages=['c1', 'c5']) 248 | # print(oupt[0].shape, oupt[1].shape) 249 | # print(model.summary()) 250 | 251 | -------------------------------------------------------------------------------- /model/pretrained/watnet.h5: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/xinluo2018/WatNet/7ff53a8e89dcf74c24392f6ba8f31d9277662122/model/pretrained/watnet.h5 -------------------------------------------------------------------------------- /model/seg_model/deeplabv3_plus.py: -------------------------------------------------------------------------------- 1 | 2 | """ 3 | @Reference: https://github.com/luyanger1799/amazing-semantic-segmentation 4 | """ 5 | import os 6 | import sys 7 | sys.path.append(os.getcwd()) 8 | import tensorflow as tf 9 | import tensorflow.keras.layers as layers 10 | import tensorflow.keras.models as models 11 | import tensorflow.keras.backend as backend 12 | from model.base_model.xception import Xception 13 | 14 | class GlobalAveragePooling2D(layers.GlobalAveragePooling2D): 15 | def __init__(self, keep_dims=False, **kwargs): 16 | super(GlobalAveragePooling2D, self).__init__(**kwargs) 17 | self.keep_dims = keep_dims 18 | def call(self, inputs): 19 | if self.keep_dims is False: 20 | return super(GlobalAveragePooling2D, self).call(inputs) 21 | else: 22 | return backend.mean(inputs, axis=[1, 2], keepdims=True) 23 | 24 | class Concatenate(layers.Concatenate): 25 | def __init__(self, out_size=None, axis=-1, name=None): 26 | super(Concatenate, self).__init__(axis=axis, name=name) 27 | self.out_size = out_size 28 | def call(self, inputs): 29 | return backend.concatenate(inputs, self.axis) 30 | 31 | def _conv_bn_relu(x, filters, kernel_size, strides=1): 32 | x = layers.Conv2D(filters, kernel_size, strides=strides, padding='same')(x) 33 | x = layers.BatchNormalization()(x) 34 | x = layers.ReLU()(x) 35 | return x 36 | 37 | def _aspp(x, out_filters, aspp_size): 38 | xs = list() 39 | x1 = layers.Conv2D(out_filters, 1, strides=1)(x) 40 | xs.append(x1) 41 | 42 | for i in range(3): 43 | xi = layers.Conv2D(out_filters, 3, strides=1, padding='same', dilation_rate=6 * (i + 1))(x) 44 | xs.append(xi) 45 | img_pool = GlobalAveragePooling2D(keep_dims=True)(x) 46 | img_pool = layers.Conv2D(out_filters, 1, 1, kernel_initializer='he_normal')(img_pool) 47 | img_pool = layers.UpSampling2D(size=aspp_size, interpolation='bilinear')(img_pool) 48 | xs.append(img_pool) 49 | 50 | x = Concatenate(out_size=aspp_size)(xs) 51 | x = layers.Conv2D(out_filters, 1, strides=1, kernel_initializer='he_normal')(x) 52 | x = layers.BatchNormalization()(x) 53 | return x 54 | 55 | def deeplabv3_plus(nclasses, input_shape=(256,256,6)): 56 | dilation = [1, 2] 57 | input = layers.Input(shape=input_shape) 58 | aspp_size = (input_shape[0] // 16, input_shape[1] // 16) 59 | encoder = Xception(version='Xception-DeepLab', dilation=dilation) 60 | c2, c5 = encoder(input, output_stages=['c1', 'c5']) 61 | x = _aspp(c5, 256, aspp_size) 62 | x = layers.Dropout(rate=0.5)(x) 63 | 64 | x = layers.UpSampling2D(size=(4, 4), interpolation='bilinear')(x) 65 | x = _conv_bn_relu(x, 48, 1, strides=1) 66 | 67 | x = Concatenate(out_size=aspp_size)([x, c2]) 68 | x = _conv_bn_relu(x, 256, 3, 1) 69 | x = layers.Dropout(rate=0.5)(x) 70 | 71 | x = _conv_bn_relu(x, 256, 3, 1) 72 | x = layers.Dropout(rate=0.1)(x) 73 | if nclasses == 2: 74 | x = layers.Conv2D(1, 1, strides=1, activation= 'sigmoid')(x) 75 | else: 76 | x = layers.Conv2D(nclasses, 1, strides=1, activation='softmax')(x) 77 | x = layers.UpSampling2D(size=(4, 4), interpolation='bilinear')(x) 78 | outputs = x 79 | # return outputs 80 | return models.Model(input, outputs, name='deeplabv3_plus') 81 | 82 | 83 | # input_img = tf.ones([4, 512, 512, 6], tf.float32) 84 | # model = deeplabv3_plus(nclasses=2, input_shape=(512,512,6)) 85 | # oupt = model(inputs=input_img) 86 | # print(oupt.shape) 87 | # model.summary() 88 | 89 | -------------------------------------------------------------------------------- /model/seg_model/deepwatermapv2.py: -------------------------------------------------------------------------------- 1 | 2 | ''' Implementation of DeepWaterMapV2. 3 | The model architecture is explained in: 4 | L.F. Isikdogan, A.C. Bovik, and P. Passalacqua, 5 | "Seeing Through the Clouds with DeepWaterMap," IEEE GRSL, 2019. 6 | ''' 7 | 8 | import tensorflow as tf 9 | 10 | def deepwatermapv2(min_width=4): 11 | inputs = tf.keras.layers.Input(shape=[None, None, 6]) 12 | 13 | def conv_block(x, num_filters, kernel_size, stride=1, use_relu=True): 14 | x = tf.keras.layers.Conv2D( 15 | filters=num_filters, 16 | kernel_size=kernel_size, 17 | kernel_initializer='he_uniform', 18 | strides=stride, 19 | padding='same', 20 | use_bias=False)(x) 21 | x = tf.keras.layers.BatchNormalization()(x) 22 | if use_relu: 23 | x = tf.keras.layers.Activation('relu')(x) 24 | return x 25 | 26 | def downscaling_unit(x): 27 | num_filters = int(x.get_shape()[-1]) * 4 28 | x_1 = conv_block(x, num_filters, kernel_size=5, stride=2) 29 | x_2 = conv_block(x_1, num_filters, kernel_size=3, stride=1) 30 | x = tf.keras.layers.Add()([x_1, x_2]) 31 | return x 32 | 33 | def upscaling_unit(x): 34 | num_filters = int(x.get_shape()[-1]) // 4 35 | x = tf.keras.layers.Lambda(lambda x: tf.nn.depth_to_space(x, 2))(x) 36 | x_1 = conv_block(x, num_filters, kernel_size=3) 37 | x_2 = conv_block(x_1, num_filters, kernel_size=3) 38 | x = tf.keras.layers.Add()([x_1, x_2]) 39 | return x 40 | 41 | def bottleneck_unit(x): 42 | num_filters = int(x.get_shape()[-1]) 43 | x_1 = conv_block(x, num_filters, kernel_size=3) 44 | x_2 = conv_block(x_1, num_filters, kernel_size=3) 45 | x = tf.keras.layers.Add()([x_1, x_2]) 46 | return x 47 | 48 | # model flow 49 | skip_connections = [] 50 | num_filters = min_width 51 | 52 | # first layer 53 | x = conv_block(inputs, num_filters, kernel_size=1, use_relu=False) 54 | skip_connections.append(x) 55 | 56 | # encoder 57 | for i in range(4): 58 | x = downscaling_unit(x) 59 | skip_connections.append(x) 60 | 61 | # bottleneck 62 | x = bottleneck_unit(x) 63 | 64 | # decoder 65 | for i in range(4): 66 | x = tf.keras.layers.Add()([x, skip_connections.pop()]) 67 | x = upscaling_unit(x) 68 | 69 | # last layer 70 | x = tf.keras.layers.Add()([x, skip_connections.pop()]) 71 | x = conv_block(x, 1, kernel_size=1, use_relu=False) 72 | x = tf.keras.layers.Activation('sigmoid')(x) 73 | 74 | model = tf.keras.Model(inputs=inputs, outputs=x) 75 | return model 76 | 77 | # model = deepwatermapv2(min_width=4) 78 | # model.summary() -------------------------------------------------------------------------------- /model/seg_model/watnet.py: -------------------------------------------------------------------------------- 1 | import os 2 | import sys 3 | sys.path.append(os.getcwd()) 4 | import tensorflow as tf 5 | import tensorflow.keras.backend as backend 6 | import tensorflow.keras.layers as layers 7 | import tensorflow.keras.models as models 8 | from model.base_model.mobilenetv2 import MobileNetV2 9 | 10 | ##### improved DeepLabV3+ 11 | def upsample(tensor, size_1): 12 | '''bilinear upsampling''' 13 | y = tf.image.resize(images=tensor, size=size_1) 14 | return y 15 | 16 | def aspp_2(tensor): 17 | '''atrous spatial pyramid pooling''' 18 | dims = backend.int_shape(tensor) 19 | y_pool = tf.keras.layers.AveragePooling2D(pool_size=( 20 | dims[1], dims[2]), name='average_pooling')(tensor) 21 | y_pool = tf.keras.layers.Conv2D(filters=128, kernel_size=1, 22 | padding='same', 23 | kernel_initializer='he_normal', 24 | name='pool_1x1conv2d', 25 | use_bias=False)(y_pool) 26 | y_pool = tf.keras.layers.BatchNormalization(name=f'bn_1')(y_pool) 27 | y_pool = tf.keras.layers.Activation('relu', name=f'relu_1')(y_pool) 28 | y_pool = upsample(tensor=y_pool, size_1=[dims[1], dims[2]]) 29 | 30 | ## 1x1 conv 31 | y_1 = tf.keras.layers.Conv2D(filters=128, kernel_size=1, 32 | dilation_rate=1, padding='same', 33 | kernel_initializer='he_normal', 34 | name='ASPP_conv2d_d1', 35 | use_bias=False)(tensor) 36 | y_1 = tf.keras.layers.BatchNormalization(name=f'bn_2')(y_1) 37 | y_1 = tf.keras.layers.Activation('relu', name=f'relu_2')(y_1) 38 | 39 | ## 3x3 dilated conv 40 | y_6 = tf.keras.layers.Conv2D(filters=128, kernel_size=3, 41 | dilation_rate=6, padding='same', 42 | kernel_initializer='he_normal', 43 | name='ASPP_conv2d_d6', 44 | use_bias=False)(tensor) 45 | y_6 = tf.keras.layers.BatchNormalization(name=f'bn_3')(y_6) 46 | y_6 = tf.keras.layers.Activation('relu', name=f'relu_3')(y_6) 47 | 48 | ## 3x3 dilated conv 49 | y_12 = tf.keras.layers.Conv2D(filters=128, kernel_size=3, 50 | dilation_rate=12, padding='same', 51 | kernel_initializer='he_normal', 52 | name='ASPP_conv2d_d12', 53 | use_bias=False)(tensor) 54 | y_12 = tf.keras.layers.BatchNormalization(name=f'bn_4')(y_12) 55 | y_12 = tf.keras.layers.Activation('relu', name=f'relu_4')(y_12) 56 | 57 | ## 3x3 dilated conv 58 | y_18 = tf.keras.layers.Conv2D(filters=128, kernel_size=3, 59 | dilation_rate=18, padding='same', 60 | kernel_initializer='he_normal', 61 | name='ASPP_conv2d_d18', 62 | use_bias=False)(tensor) 63 | y_18 = tf.keras.layers.BatchNormalization(name=f'bn_5')(y_18) 64 | y_18 = tf.keras.layers.Activation('relu', name=f'relu_5')(y_18) 65 | 66 | ## concat 67 | y = tf.keras.layers.concatenate([y_pool, y_1, y_6, y_12, y_18], name='ASPP_concat') 68 | y = tf.keras.layers.Conv2D(filters=128, kernel_size=1, 69 | dilation_rate=1, padding='same', 70 | kernel_initializer='he_normal', 71 | name='ASPP_conv2d_final', 72 | use_bias=False)(y) 73 | y = tf.keras.layers.BatchNormalization(name=f'bn_final')(y) 74 | y = tf.keras.layers.Activation('relu', name=f'relu_final')(y) 75 | return y 76 | 77 | def watnet(input_shape, nclasses=2): 78 | ''' 79 | Arguments: 80 | input_shape: (img_height, img_width, img_channel) 81 | base_model: backbone network 82 | d_feature, m_feature, l_feature: features corresponding 83 | to the deep, middle, and low layers of the backbone model 84 | nclass: number of classes. 85 | ''' 86 | print('*** Building watnet network ***') 87 | d_feature, m_feature, l_feature = 91, 24, 11 88 | (img_height, img_width, img_channel) = input_shape 89 | ## deep features 90 | base_model = MobileNetV2(input_shape, nclasses) 91 | image_features = base_model.get_layer(index = d_feature).output 92 | x_a = aspp_2(image_features) 93 | x_a = upsample(tensor=x_a, size_1=[img_height // 4, img_width // 4]) 94 | ## middle features (1/4 patch size) 95 | x_b = base_model.get_layer(index = m_feature).output 96 | x_b = layers.Conv2D(filters=48, kernel_size=1, padding='same', 97 | kernel_initializer='he_normal', name='low_level_projection', use_bias=False)(x_b) 98 | x_b = layers.BatchNormalization(name=f'bn_low_level_projection')(x_b) 99 | x_b = layers.Activation('relu', name='low_level_activation')(x_b) 100 | ## middle features (1/2 patch size) 101 | x_c = base_model.get_layer(index = l_feature).output 102 | x_c = layers.Conv2D(filters=48, 103 | kernel_size=1, 104 | padding='same', 105 | kernel_initializer='he_normal', 106 | name='low_level_projection_2', 107 | use_bias=False)(x_c) 108 | x_c = layers.BatchNormalization(name=f'bn_low_level_projection_2')(x_c) 109 | x_c = layers.Activation('relu', name='low_level_activation_2')(x_c) 110 | ## concat 111 | x = layers.concatenate([x_a, x_b], name='decoder_concat_1') 112 | x = layers.Conv2D(filters=128, 113 | kernel_size=3, 114 | padding='same', 115 | activation='relu', 116 | kernel_initializer='he_normal', 117 | name='decoder_conv2d_1', 118 | use_bias=False)(x) 119 | x = layers.BatchNormalization(name=f'bn_decoder_1')(x) 120 | x = layers.Activation('relu', name='activation_decoder_1')(x) 121 | x = layers.Conv2D(filters=128, 122 | kernel_size=3, 123 | padding='same', 124 | activation='relu', 125 | kernel_initializer='he_normal', 126 | name='decoder_conv2d_2', 127 | use_bias=False)(x) 128 | x = layers.BatchNormalization(name=f'bn_decoder_2')(x) 129 | x = layers.Activation('relu', name='activation_decoder_2')(x) 130 | x = upsample(x, [img_height//2, img_width//2]) 131 | ## concat 132 | x_2 = layers.concatenate([x, x_c], name='decoder_concat_3') 133 | x_2 = layers.Conv2DTranspose(filters=128, 134 | kernel_size=3, 135 | strides=2, 136 | padding='same', 137 | kernel_initializer='he_normal', 138 | name='decoder_deconv2d', use_bias=False)(x_2) 139 | x_2 = layers.BatchNormalization(name=f'bn_decoder_4')(x_2) 140 | x_2 = layers.Activation('relu', name='activation_decoder_4')(x_2) 141 | last = tf.keras.layers.Conv2D(1, (1,1), 142 | strides=1, 143 | padding='same', 144 | kernel_initializer='he_normal', 145 | activation= 'sigmoid') ## (bs, 256, 256, 1) 146 | x_2 = last(x_2) 147 | model = models.Model(inputs=base_model.input, outputs=x_2, name='watnet') 148 | print(f'*** Output_Shape => {model.output_shape} ***') 149 | return model 150 | 151 | # model = watnet(input_shape=(512, 512, 6), nclasses=2) 152 | # model.summary() 153 | 154 | -------------------------------------------------------------------------------- /notebooks/config.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | from utils.acc_patch import miou_binary 3 | import math 4 | ## ---- root dir ---- ## 5 | root = '/home/yons/Desktop/developer-luo/WatNet' # local sever 6 | # root = '/content/drive/My Drive/WatNet' # colab 7 | 8 | ## ---- super-parameter for model training ---- ## 9 | patch_size = 512 10 | num_bands = 6 11 | epochs = 200 12 | lr = 0.002 13 | batch_size = 4 14 | buffer_size = 200 15 | # size_tra_scene = 64 16 | size_scene = 95 17 | step_per_epoch = math.ceil(size_scene/batch_size) 18 | 19 | 20 | ## ---- configuration for model training ---- ## 21 | class lr_schedule(tf.keras.optimizers.schedules.LearningRateSchedule): 22 | def __init__(self, initial_learning_rate, steps_all): 23 | self.initial_learning_rate = initial_learning_rate 24 | self.steps_all = steps_all 25 | def __call__(self, step): 26 | return self.initial_learning_rate*((1-step/self.steps_all)**0.9) 27 | # lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay( 28 | # initial_learning_rate=lr, 29 | # decay_steps=100, # 1 step = 1 batch data 30 | # decay_rate=0.9) 31 | loss_bce = tf.keras.losses.BinaryCrossentropy() 32 | opt_adam = tf.keras.optimizers.Adam(learning_rate=\ 33 | lr_schedule(lr,step_per_epoch*epochs)) 34 | 35 | ## ---- metrics ---- ## 36 | tra_loss = tf.keras.metrics.Mean(name="tra_loss") 37 | tra_oa = tf.keras.metrics.BinaryAccuracy('tra_oa') 38 | tra_miou = miou_binary(num_classes=2, name='tra_miou') 39 | val_loss = tf.keras.metrics.Mean(name="test_loss") 40 | val_oa = tf.keras.metrics.BinaryAccuracy('test_oa') 41 | val_miou = miou_binary(num_classes=2, name='test_miou') 42 | 43 | 44 | -------------------------------------------------------------------------------- /utils/acc_patch.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | 3 | class miou_binary(tf.keras.metrics.MeanIoU): 4 | def update_state(self, y_true, y_pred, sample_weight=None): 5 | y_pred = tf.where(y_pred>0.5, 1, 0) 6 | super().update_state(y_true, y_pred, sample_weight) 7 | 8 | -------------------------------------------------------------------------------- /utils/acc_pixel.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from sklearn.metrics import accuracy_score, confusion_matrix 3 | 4 | def acc_matrix(cla_map, sam_pixel, id_label): 5 | ''' 6 | Arguments: 7 | cla_map: classification result of the full image 8 | sam_pixel: array(num_samples,3), col 1,2,3 are the row,col and label. 9 | id_label: 0,1,2... 10 | Return: 11 | the overall accuracy and confusion matrix 12 | ''' 13 | sam_result = [] 14 | num_cla = sam_pixel[:,2].max()+1 15 | labels = list(range(num_cla)) 16 | for i in range(sam_pixel.shape[0]): 17 | sam_result.append(cla_map[sam_pixel[i,0], sam_pixel[i,1]]) 18 | sam_result = np.array(sam_result) 19 | acc_oa = np.around(accuracy_score(sam_pixel[:,2], sam_result), 4) 20 | confus_mat = confusion_matrix(sam_pixel[:,2], sam_result, labels=labels) 21 | acc_prod=np.around(confus_mat[id_label,id_label]/confus_mat[id_label,:].sum(), 4) 22 | acc_user=np.around(confus_mat[id_label,id_label]/confus_mat[:,id_label].sum(), 4) 23 | 24 | return acc_oa, acc_prod, acc_user, confus_mat 25 | 26 | -------------------------------------------------------------------------------- /utils/geotif_io.py: -------------------------------------------------------------------------------- 1 | ## author: luo xin, creat: 2021.6.18, modify: 2021.7.14 2 | 3 | import numpy as np 4 | from osgeo import gdal 5 | from osgeo import osr 6 | 7 | ### tiff image reading 8 | def readTiff(path_in): 9 | ''' 10 | return: 11 | img: numpy array, exent: tuple, (x_min, x_max, y_min, y_max) 12 | proj info, and dimentions: (row, col, band) 13 | ''' 14 | RS_Data=gdal.Open(path_in) 15 | im_col = RS_Data.RasterXSize # 16 | im_row = RS_Data.RasterYSize # 17 | im_bands =RS_Data.RasterCount # 18 | im_geotrans = RS_Data.GetGeoTransform() # 19 | im_proj = RS_Data.GetProjection() # 20 | img_array = RS_Data.ReadAsArray(0, 0, im_col, im_row).astype(np.float) # 21 | left = im_geotrans[0] 22 | up = im_geotrans[3] 23 | right = left + im_geotrans[1] * im_col + im_geotrans[2] * im_row 24 | bottom = up + im_geotrans[5] * im_row + im_geotrans[4] * im_col 25 | extent = (left, right, bottom, up) 26 | espg_code = osr.SpatialReference(wkt=im_proj).GetAttrValue('AUTHORITY',1) 27 | 28 | img_info = {'geoextent': extent, 'geotrans':im_geotrans, \ 29 | 'geosrs': espg_code, 'row': im_row, 'col': im_col,\ 30 | 'bands': im_bands} 31 | 32 | if im_bands > 1: 33 | img_array = np.transpose(img_array, (1, 2, 0)) # 34 | return img_array, img_info 35 | else: 36 | return img_array, img_info 37 | 38 | ### .tiff image write 39 | def writeTiff(im_data, im_geotrans, im_geosrs, path_out): 40 | ''' 41 | input: 42 | im_data: tow dimentions (order: row, col),or three dimentions (order: row, col, band) 43 | im_geosrs: espg code correspond to image spatial reference system. 44 | ''' 45 | im_data = np.squeeze(im_data) 46 | if 'int8' in im_data.dtype.name: 47 | datatype = gdal.GDT_Byte 48 | elif 'int16' in im_data.dtype.name: 49 | datatype = gdal.GDT_UInt16 50 | else: 51 | datatype = gdal.GDT_Float32 52 | if len(im_data.shape) == 3: 53 | im_data = np.transpose(im_data, (2, 0, 1)) 54 | im_bands, im_height, im_width = im_data.shape 55 | else: 56 | im_bands,(im_height, im_width) = 1,im_data.shape 57 | driver = gdal.GetDriverByName("GTiff") 58 | dataset = driver.Create(path_out, im_width, im_height, im_bands, datatype) 59 | if(dataset!= None): 60 | dataset.SetGeoTransform(im_geotrans) # 61 | dataset.SetProjection("EPSG:" + str(im_geosrs)) # 62 | if im_bands > 1: 63 | for i in range(im_bands): 64 | dataset.GetRasterBand(i+1).WriteArray(im_data[i]) 65 | del dataset 66 | else: 67 | dataset.GetRasterBand(1).WriteArray(im_data) 68 | del dataset -------------------------------------------------------------------------------- /utils/imgPatch.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | 3 | class imgPatch(): 4 | ''' 5 | author: xin luo, date: 2021.3.19 6 | description: 1. remote sensing image to multi-scale patches 7 | 2. patches to remote sensing image 8 | ''' 9 | def __init__(self, img, patch_size, edge_overlay): 10 | ''' edge_overlay = left overlay or, right overlay 11 | edge_overlay should be an even number. ''' 12 | self.patch_size = patch_size 13 | self.edge_overlay = edge_overlay 14 | self.img = img[:,:,np.newaxis] if len(img.shape) == 2 else img 15 | self.img_row = img.shape[0] 16 | self.img_col = img.shape[1] 17 | 18 | def toPatch(self): 19 | ''' 20 | description: convert img to patches. 21 | return: 22 | patch_list, contains all generated patches. 23 | start_list, contains all start positions(row, col) of the generated patches. 24 | ''' 25 | patch_list = [] 26 | start_list = [] 27 | patch_step = self.patch_size - self.edge_overlay 28 | img_expand = np.pad(self.img, ((self.edge_overlay, patch_step), 29 | (self.edge_overlay, patch_step), (0,0)), 'constant') 30 | img_patch_row = (img_expand.shape[0]-self.edge_overlay)//patch_step 31 | img_patch_col = (img_expand.shape[1]-self.edge_overlay)//patch_step 32 | for i in range(img_patch_row): 33 | for j in range(img_patch_col): 34 | patch_list.append(img_expand[i*patch_step:i*patch_step+self.patch_size, 35 | j*patch_step:j*patch_step+self.patch_size, :]) 36 | start_list.append([i*patch_step-self.edge_overlay, j*patch_step-self.edge_overlay]) 37 | return patch_list, start_list, img_patch_row, img_patch_col 38 | 39 | def higher_patch_crop(self, higher_patch_size, start_list): 40 | ''' 41 | author: xin luo, date: 2021.3.19 42 | description: crop the higher-scale patch (centered by the given low-scale patch) 43 | input: 44 | img, np.array, the original image 45 | patch_size, int, the lower-scale patch size 46 | crop_size, int, the higher-scale patch size 47 | start_list, list, the start position (row,col) corresponding to the original image (generated by the toPatch function) 48 | return: 49 | higher_patch_list, list, contains higher-scale patches corresponding to the lower-scale patches. 50 | ''' 51 | higher_patch_list = [] 52 | radius_bias = higher_patch_size//2-self.patch_size//2 53 | patch_step = self.patch_size - self.edge_overlay 54 | img_expand = np.pad(self.img, ((self.edge_overlay, patch_step), (self.edge_overlay, patch_step), (0,0)), 'constant') 55 | img_expand_higher = np.pad(img_expand, ((radius_bias, radius_bias), (radius_bias, radius_bias), (0,0)), 'constant') 56 | start_list_new = list(np.array(start_list)+self.edge_overlay+radius_bias) 57 | for start_i in start_list_new: 58 | higher_row_start, higher_col_start = start_i[0]-radius_bias, start_i[1]-radius_bias 59 | higher_patch = img_expand_higher[higher_row_start:higher_row_start+higher_patch_size,higher_col_start:higher_col_start+higher_patch_size,:] 60 | higher_patch_list.append(higher_patch) 61 | return higher_patch_list 62 | 63 | def toImage(self, patch_list, img_patch_row, img_patch_col): 64 | patch_list = [patch[self.edge_overlay//2:-self.edge_overlay//2, self.edge_overlay//2:-self.edge_overlay//2,:] 65 | for patch in patch_list] 66 | patch_list = [np.hstack((patch_list[i*img_patch_col:i*img_patch_col+img_patch_col])) 67 | for i in range(img_patch_row)] 68 | img_array = np.vstack(patch_list) 69 | img_array = img_array[self.edge_overlay//2:self.img_row+self.edge_overlay//2, \ 70 | self.edge_overlay//2:self.img_col+self.edge_overlay//2,:] 71 | 72 | return img_array -------------------------------------------------------------------------------- /utils/imgShow.py: -------------------------------------------------------------------------------- 1 | ## date: 2021.6.2 2 | 3 | import matplotlib.pyplot as plt 4 | import numpy as np 5 | 6 | def imgShow(img, extent=None, color_bands=(2,1,0), \ 7 | clip_percent=2, per_band_clip='False'): 8 | ''' 9 | Arguments: 10 | img: (row, col, band) or (row, col) 11 | num_bands: a list/tuple, [red_band,green_band,blue_band] 12 | clip_percent: for linear strech, value within the range of 0-100. 13 | per_band: if 'True', the band values will be clipped by each band respectively. 14 | ''' 15 | img = img/(np.amax(img)+0.00001) # normalization 16 | img = np.squeeze(img) 17 | while np.isnan(np.sum(img)) == True: 18 | where_are_NaNs = np.isnan(img) 19 | img[where_are_NaNs] = 0 20 | if np.min(img) == np.max(img): 21 | if len(img.shape) == 2: 22 | plt.imshow(np.clip(img, 0, 1), extent=extent, vmin=0,vmax=1) 23 | else: 24 | plt.imshow(np.clip(img[:,:,0], 0, 1), extent=extent, vmin=0,vmax=1) 25 | else: 26 | if len(img.shape) == 2: 27 | img_color = np.expand_dims(img, axis=2) 28 | else: 29 | img_color = img[:,:,[color_bands[0], color_bands[1], color_bands[2]]] 30 | img_color_clip = np.zeros_like(img_color) 31 | if per_band_clip == 'True': 32 | for i in range(img_color.shape[-1]): 33 | img_color_hist = np.percentile(img_color[:,:,i], [clip_percent, 100-clip_percent]) 34 | img_color_clip[:,:,i] = (img_color[:,:,i]-img_color_hist[0])\ 35 | /(img_color_hist[1]-img_color_hist[0]+0.0001) 36 | else: 37 | img_color_hist = np.percentile(img_color, [clip_percent, 100-clip_percent]) 38 | img_color_clip = (img_color-img_color_hist[0])\ 39 | /(img_color_hist[1]-img_color_hist[0]+0.0001) 40 | 41 | img_color_clip = np.squeeze(img_color_clip) 42 | plt.imshow(np.clip(img_color_clip, 0, 1), extent=extent, vmin=0,vmax=1) 43 | 44 | 45 | def imsShow(img_list, img_name_list, clip_list=None, color_bands_list=None): 46 | ''' des: visualize multiple images. 47 | input: 48 | img_list: containes all images 49 | img_names_list: image names corresponding to the images 50 | clip_list: percent clips (histogram) corresponding to the images 51 | color_bands_list: color bands combination corresponding to the images 52 | ''' 53 | if not clip_list: 54 | clip_list = [0 for i in range(len(img_list))] 55 | if not color_bands_list: 56 | color_bands_list = [[2, 1, 0] for i in range(len(img_list))] 57 | for i in range(len(img_list)): 58 | plt.subplot(1, len(img_list), i+1) 59 | plt.title(img_name_list[i]) 60 | imgShow(img=img_list[i],\ 61 | color_bands=color_bands_list[i], clip_percent=clip_list[i]) 62 | plt.axis('off') 63 | -------------------------------------------------------------------------------- /watnet_infer.py: -------------------------------------------------------------------------------- 1 | ## author: xin luo, creat: 2021.8.11 2 | 3 | ''' 4 | des: perform surface water mapping by using pretrained watnet 5 | through funtional api and command line, respectively. 6 | 7 | example: 8 | funtional api: 9 | water_map = watnet_infer(rsimg) 10 | command line: 11 | python watnet_infer.py data/test-demo/*.tif 12 | python watnet_infer.py data/test-demo/*.tif -o data/test-demo/result 13 | note: 14 | rsimg is np.array (row,col,band), value: [0,1] 15 | data/test-demo/*.tif is the sentinel-2 image path 16 | data/test-demo/result is output directory 17 | ''' 18 | 19 | import os 20 | import numpy as np 21 | import tensorflow as tf 22 | import argparse 23 | from utils.imgPatch import imgPatch 24 | from utils.geotif_io import readTiff,writeTiff 25 | 26 | ## default path of the pretrained watnet model 27 | path_watnet = 'model/pretrained/watnet.h5' 28 | 29 | def get_args(): 30 | 31 | description = 'surface water mapping by using pretrained watnet' 32 | parser = argparse.ArgumentParser(description=description) 33 | 34 | parser.add_argument( 35 | 'ifile', metavar='ifile', type=str, nargs='+', 36 | help=('file(s) to process (.tiff)')) 37 | 38 | parser.add_argument( 39 | '-m', metavar='watnet', dest='watnet', type=str, 40 | nargs='+', default=path_watnet, 41 | help=('pretrained watnet model (tensorflow2, .h5)')) 42 | 43 | parser.add_argument( 44 | '-o', metavar='odir', dest='odir', type=str, nargs='+', 45 | help=('directory to write')) 46 | 47 | return parser.parse_args() 48 | 49 | 50 | def watnet_infer(rsimg, path_model = path_watnet): 51 | 52 | ''' des: surface water mapping by using pretrained watnet 53 | arg: 54 | img: np.array, surface reflectance data (!!data value: 0-1), 55 | consist of 6 bands (blue,green,red,nir,swir-1,swir-2). 56 | path_model: str, the path of the pretrained model. 57 | retrun: 58 | water_map: np.array. 59 | ''' 60 | ### ----- load the pretrained model -----# 61 | model = tf.keras.models.load_model(path_model, compile=False) 62 | ### ------ apply the pre-trained model 63 | imgPatch_ins = imgPatch(rsimg, patch_size=512, edge_overlay=80) 64 | patch_list, start_list, img_patch_row, img_patch_col = imgPatch_ins.toPatch() 65 | result_patch_list = [model(patch[np.newaxis, :]) for patch in patch_list] 66 | result_patch_list = [np.squeeze(patch, axis = 0) for patch in result_patch_list] 67 | pro_map = imgPatch_ins.toImage(result_patch_list, img_patch_row, img_patch_col) 68 | water_map = np.where(pro_map>0.5, 1, 0) 69 | 70 | return water_map 71 | 72 | 73 | if __name__ == '__main__': 74 | args = get_args() 75 | ifile = args.ifile 76 | path_model = args.watnet 77 | odir = args.odir 78 | ## write path 79 | if odir: 80 | if not os.path.exists(odir[0]): 81 | os.makedirs(odir[0]) 82 | ofile = [os.path.splitext(file)[0] + '_water.tif' for file in ifile] 83 | ofile = [os.path.join(odir[0], os.path.split(file)[1]) for file in ofile] 84 | else: 85 | ofile = [os.path.splitext(file)[0] + '_water.tif' for file in ifile] 86 | 87 | for i in range(len(ifile)): 88 | print('file in -->', ifile[i]) 89 | ## image reading and normalization 90 | sen2_img, img_info = readTiff(path_in=ifile[i]) 91 | sen2_img = np.float32(np.clip(sen2_img/10000, a_min=0, a_max=1)) ## normalization 92 | ## surface water mapping by using watnet 93 | water_map = watnet_infer(rsimg=sen2_img) 94 | # write out the result 95 | print('write out -->', ofile[i]) 96 | writeTiff(im_data = water_map.astype(np.int8), 97 | im_geotrans = img_info['geotrans'], 98 | im_geosrs = img_info['geosrs'], 99 | path_out = ofile[i]) 100 | 101 | 102 | --------------------------------------------------------------------------------