Satellite imagery

Introduction
This guide addresses a core topic in remote sensing: the automatic pixelwise labeling of aerial imagery. There are many use cases for analytics on satellite data ranging from tracking of different objects to segmenting houses, roads, etc. Today Deep Learning approaches are at the core of many production solutions. When applying machine learning (ML) to satellite data there are a few typical challenges that often arise:
The size of images: Images may have high resolution, so data will be in the tens of gigabytes. With continuous data updates image collection can quickly reach multiple terabytes.
Custom machine learning: many out-of-the-box solutions do not achive acceptable accuracies without customization. In addition sometimes unconventional pre- and post- processing steps are required.
Task description
Today we are going to demonstrate how users can build a custom neural network for buildings segmentation inside Supervisely.
We will show how to perform special training data preparations (DTL layer: sliding window crops), train a neural network, apply it to test images in a sliding window manner and then postprocess predictions (split single prediction mask into separate objects).
For this tutorial we will use The Inria Aerial Image Labeling Dataset.
You can download the data and reproduce the entire research yourself. The data may only be used for non-commercial purposes.
Step 1. Upload INRIA dataset
1.
Download original dataset​
2.
Unzip the archive, open the train folder and rename the gt folder to ann
3.
Choose the binary_masks import option
4.
Drag images and ann dirs to the upload window
5.
Name this project INRIA
Step 2. DTL #1 to split data to train/val/test
This DTL query splits the original dataset into train-validation and test subsets taking approximately the same number of images from each town (around 5 images per town).
1.
Layer #1 ("action": "data") takes all data from project INRIA and keeps classes as they are. Then there are five indentical branches for each town (Austin, Chicago, Kitsap County, Vienna and West Tyrol), so we will describe only one of them for Austin.
2.
Layer #2 ("action": "if") selects the data related to austin, other data is sent to $null1.
3.
Layer #3 ("action": "tag") adds tag austin to each input image.
4.
Layer #4 ("action": "if") randomly splits the data into two branches: first branch - 14% (will be tagged as trainval) and second branch - 86% (will be tagged as test). We decreased the number of training images to demonstrate that entire pipline can be passed with a small number of annotated images.
5.
Layer #5 ("action": "tag") adds tag train to all input images.
6.
Layer #6 ("action": "tag") adds tag val to all input images .
7.
Layer #7 ("action": "supervisely") saves results to the new project INRIA_TAGGED.
1
[
2
  {
3
    "dst": "$sample",
4
    "src": [
5
      "INRIA/*"
6
    ],
7
    "action": "data",
8
    "settings": {
9
      "classes_mapping": "default"
10
    }
11
  },
12
  {
13
    "dst": [
14
      "$1",
15
      "null"
16
    ],
17
    "src": [
18
      "$sample"
19
    ],
20
    "action": "if",
21
    "settings": {
22
      "condition": {
23
        "regex_names": [
24
          "^austin[0-9]+quot;
25
        ]
26
      }
27
    }
28
  },
29
  {
30
    "dst": "$tag_1",
31
    "src": [
32
      "$1"
33
    ],
34
    "action": "tag",
35
    "settings": {
36
      "tag": "austin",
37
      "action": "add"
38
    }
39
  },
40
  {
41
    "dst": [
42
      "$to_train1",
43
      "$to_val1"
44
    ],
45
    "src": [
46
      "$tag_1"
47
    ],
48
    "action": "if",
49
    "settings": {
50
      "condition": {
51
        "probability": 0.14
52
      }
53
    }
54
  },
55
  {
56
    "dst": "$train1",
57
    "src": [
58
      "$to_train1"
59
    ],
60
    "action": "tag",
61
    "settings": {
62
      "tag": "trainval",
63
      "action": "add"
64
    }
65
  },
66
  {
67
    "dst": "$val1",
68
    "src": [
69
      "$to_val1"
70
    ],
71
    "action": "tag",
72
    "settings": {
73
      "tag": "test",
74
      "action": "add"
75
    }
76
  },
77
  {
78
    "dst": [
79
      "$2",
80
      "null"
81
    ],
82
    "src": [
83
      "$sample"
84
    ],
85
    "action": "if",
86
    "settings": {
87
      "condition": {
88
        "regex_names": [
89
          "^chicago[0-9]+quot;
90
        ]
91
      }
92
    }
93
  },
94
  {
95
    "dst": "$tag_2",
96
    "src": [
97
      "$2"
98
    ],
99
    "action": "tag",
100
    "settings": {
101
      "tag": "chicago",
102
      "action": "add"
103
    }
104
  },
105
  {
106
    "dst": [
107
      "$to_train2",
108
      "$to_val2"
109
    ],
110
    "src": [
111
      "$tag_2"
112
    ],
113
    "action": "if",
114
    "settings": {
115
      "condition": {
116
        "probability": 0.14
117
      }
118
    }
119
  },
120
  {
121
    "dst": "$train2",
122
    "src": [
123
      "$to_train2"
124
    ],
125
    "action": "tag",
126
    "settings": {
127
      "tag": "trainval",
128
      "action": "add"
129
    }
130
  },
131
  {
132
    "dst": "$val2",
133
    "src": [
134
      "$to_val2"
135
    ],
136
    "action": "tag",
137
    "settings": {
138
      "tag": "test",
139
      "action": "add"
140
    }
141
  },
142
  {
143
    "dst": [
144
      "$3",
145
      "null"
146
    ],
147
    "src": [
148
      "$sample"
149
    ],
150
    "action": "if",
151
    "settings": {
152
      "condition": {
153
        "regex_names": [
154
          "^kitsap[0-9]+quot;
155
        ]
156
      }
157
    }
158
  },
159
  {
160
    "dst": "$tag_3",
161
    "src": [
162
      "$3"
163
    ],
164
    "action": "tag",
165
    "settings": {
166
      "tag": "kitsap",
167
      "action": "add"
168
    }
169
  },
170
  {
171
    "dst": [
172
      "$to_train3",
173
      "$to_val3"
174
    ],
175
    "src": [
176
      "$tag_3"
177
    ],
178
    "action": "if",
179
    "settings": {
180
      "condition": {
181
        "probability": 0.14
182
      }
183
    }
184
  },
185
  {
186
    "dst": "$train3",
187
    "src": [
188
      "$to_train3"
189
    ],
190
    "action": "tag",
191
    "settings": {
192
      "tag": "trainval",
193
      "action": "add"
194
    }
195
  },
196
  {
197
    "dst": "$val3",
198
    "src": [
199
      "$to_val3"
200
    ],
201
    "action": "tag",
202
    "settings": {
203
      "tag": "test",
204
      "action": "add"
205
    }
206
  },
207
  {
208
    "dst": [
209
      "$4",
210
      "null"
211
    ],
212
    "src": [
213
      "$sample"
214
    ],
215
    "action": "if",
216
    "settings": {
217
      "condition": {
218
        "regex_names": [
219
          "^tyrol-w[0-9]+quot;
220
        ]
221
      }
222
    }
223
  },
224
  {
225
    "dst": "$tag_4",
226
    "src": [
227
      "$4"
228
    ],
229
    "action": "tag",
230
    "settings": {
231
      "tag": "tyrol-w",
232
      "action": "add"
233
    }
234
  },
235
  {
236
    "dst": [
237
      "$to_train4",
238
      "$to_val4"
239
    ],
240
    "src": [
241
      "$tag_4"
242
    ],
243
    "action": "if",
244
    "settings": {
245
      "condition": {
246
        "probability": 0.14
247
      }
248
    }
249
  },
250
  {
251
    "dst": "$train4",
252
    "src": [
253
      "$to_train4"
254
    ],
255
    "action": "tag",
256
    "settings": {
257
      "tag": "trainval",
258
      "action": "add"
259
    }
260
  },
261
  {
262
    "dst": "$val4",
263
    "src": [
264
      "$to_val4"
265
    ],
266
    "action": "tag",
267
    "settings": {
268
      "tag": "test",
269
      "action": "add"
270
    }
271
  },
272
  {
273
    "dst": [
274
      "$5",
275
      "null"
276
    ],
277
    "src": [
278
      "$sample"
279
    ],
280
    "action": "if",
281
    "settings": {
282
      "condition": {
283
        "regex_names": [
284
          "^vienna[0-9]+quot;
285
        ]
286
      }
287
    }
288
  },
289
  {
290
    "dst": "$tag_5",
291
    "src": [
292
      "$5"
293
    ],
294
    "action": "tag",
295
    "settings": {
296
      "tag": "vienna",
297
      "action": "add"
298
    }
299
  },
300
  {
301
    "dst": [
302
      "$to_train5",
303
      "$to_val5"
304
    ],
305
    "src": [
306
      "$tag_5"
307
    ],
308
    "action": "if",
309
    "settings": {
310
      "condition": {
311
        "probability": 0.14
312
      }
313
    }
314
  },
315
  {
316
    "dst": "$train5",
317
    "src": [
318
      "$to_train5"
319
    ],
320
    "action": "tag",
321
    "settings": {
322
      "tag": "trainval",
323
      "action": "add"
324
    }
325
  },
326
  {
327
    "dst": "$val5",
328
    "src": [
329
      "$to_val5"
330
    ],
331
    "action": "tag",
332
    "settings": {
333
      "tag": "test",
334
      "action": "add"
335
    }
336
  },
337
  {
338
    "dst": "INRIA_TAGGED",
339
    "src": [
340
      "$train1",
341
      "$val1",
342
      "$train2",
343
      "$val2",
344
      "$train3",
345
      "$val3",
346
      "$train4",
347
      "$val4",
348
      "$train5",
349
      "$val5"
350
    ],
351
    "action": "supervisely",
352
    "settings": {}
353
  }
354
]
Copied!
Step 3. DTL #2 prepare training project
In this DTL query we filter images, apply transformations and split them into train and validation sets.
1.
Layer #1 ("action": "data") takes all data from project INRIA_TAGGED and keeps classes as they are.
2.
Layer #2 ("action": "if") splits the data into two branches based on tags. Further we will work only with the images which have tag trainval, other images will be sent to null.
3.
Layer #3 ("action": "sliding_window") creates crops of size 512x512 from the images using sliding window approach.
4.
Layer #4 ("action": "flip") flips data horisontally.
5.
Layer #5 ("action": "flip") flips data vertically.
6.
Layer #6 ("action": "if") randomly splits the data into two branches: first branch - 95% (will be tagged as train) and second branch - 5% (will be tagged as val).
7.
Layer #7 ("action": "tag") adds the tag train to all input images .
8.
Layer #8 ("action": "tag") adds the tag val to all input images.
9.
Layer #9 ("action": "if") filters the data (images without objects will be skipped).
10.
Layer #10 ("action": "supervisely") saves results to the new project INRIA_train. As a result we got around 7000 images for training.
1
[
2
  {
3
    "dst": "$sample",
4
    "src": [
5
      "INRIA_TAGGED/*"
6
    ],
7
    "action": "data",
8
    "settings": {
9
      "classes_mapping": "default"
10
    }
11
  },
12
  {
13
    "dst": [
14
      "$sample1",
15
      "null"
16
    ],
17
    "src": [
18
      "$sample"
19
    ],
20
    "action": "if",
21
    "settings": {
22
      "condition": {
23
        "tags": [
24
          "trainval"
25
        ]
26
      }
27
    }
28
  },
29
  {
30
    "dst": "$crops",
31
    "src": [
32
      "$sample1"
33
    ],
34
    "action": "sliding_window",
35
    "settings": {
36
      "window": {
37
        "width": 512,
38
        "height": 512
39
      },
40
      "min_overlap": {
41
        "x": 0,
42
        "y": 0
43
      }
44
    }
45
  },
46
  {
47
    "dst": "$vflip",
48
    "src": [
49
      "$crops"
50
    ],
51
    "action": "flip",
52
    "settings": {
53
      "axis": "vertical"
54
    }
55
  },
56
  {
57
    "dst": "$hflip",
58
    "src": [
59
      "$crops",
60
      "$vflip"
61
    ],
62
    "action": "flip",
63
    "settings": {
64
      "axis": "horizontal"
65
    }
66
  },
67
  {
68
    "dst": [
69
      "$to_train",
70
      "$to_val"
71
    ],
72
    "src": [
73
      "$crops",
74
      "$hflip",
75
      "$vflip"
76
    ],
77
    "action": "if",
78
    "settings": {
79
      "condition": {
80
        "probability": 0.95
81
      }
82
    }
83
  },
84
  {
85
    "dst": "$train",
86
    "src": [
87
      "$to_train"
88
    ],
89
    "action": "tag",
90
    "settings": {
91
      "tag": "train",
92
      "action": "add"
93
    }
94
  },
95
  {
96
    "dst": "$val",
97
    "src": [
98
      "$to_val"
99
    ],
100
    "action": "tag",
101
    "settings": {
102
      "tag": "val",
103
      "action": "add"
104
    }
105
  },
106
  {
107
    "dst": [
108
      "$res",
109
      "null"
110
    ],
111
    "src": [
112
      "$train",
113
      "$val"
114
    ],
115
    "action": "if",
116
    "settings": {
117
      "condition": {
118
        "min_objects_count": 1
119
      }
120
    }
121
  },
122
  {
123
    "dst": "INRIA_train",
124
    "src": [
125
      "$res"
126
    ],
127
    "action": "supervisely",
128
    "settings": {}
129
  }
130
]
Copied!
Step 3. UNet training
Basic step by step training guide is here. It is the same for all models inside Supervisely. Detailed information regarding training configs is here.
UNetV2's encoder weights were initialized from corresponding model from Model Zoo (UNetV2 with VGG weigths that was pretrained on ImageNet).
Training config:
1
{
2
  "lr": 0.001,
3
  "epochs": 5,
4
  "val_every": 1,
5
  "batch_size": {
6
    "val": 3,
7
    "train": 3
8
  },
9
  "input_size": {
10
    "width": 512,
11
    "height": 512
12
  },
13
  "gpu_devices": [
14
    0,
15
    1,
16
    2
17
  ],
18
  "data_workers": {
19
    "val": 3,
20
    "train": 3
21
  },
22
  "dataset_tags": {
23
    "val": "val",
24
    "train": "train"
25
  },
26
  "special_classes": {
27
    "neutral": "neutral",
28
    "background": "bg"
29
  },
30
  "weights_init_type": "transfer_learning"
31
}
Copied!
Training took ~ 1 hour.
Here are the training charts:
Step 4. Inference
Sliding window inference:
1
{
2
  "mode": {
3
    "save": false,
4
    "source": "sliding_window",
5
    "window": {
6
      "width": 512,
7
      "height": 512
8
    },
9
    "min_overlap": {
10
      "x": 128,
11
      "y": 128
12
    }
13
  },
14
  "gpu_devices": [
15
    0
16
  ],
17
  "model_classes": {
18
    "add_suffix": "_dl",
19
    "save_classes": "__all__"
20
  },
21
  "existing_objects": {
22
    "add_suffix": "",
23
    "save_classes": []
24
  }
25
}
Copied!
Here is an example of NN predictions:
Step 5. DTL #3 to split prediction mask to separate objects
Before splitting all buildings are on a single mask:
After the DTL query all buildings are split into different masks:
Here is a simple example of the DTL query:
1
[
2
  {
3
    "dst": "$sample",
4
    "src": [
5
      "inf_INRIA_test_part/*"
6
    ],
7
    "action": "data",
8
    "settings": {
9
      "classes_mapping": "default"
10
    }
11
  },
12
  {
13
    "dst": "$splitted",
14
    "src": [
15
      "$sample"
16
    ],
17
    "action": "split_masks",
18
    "settings": {
19
      "classes": [
20
        "untitled_dl"
21
      ]
22
    }
23
  },
24
  {
25
    "dst": "inf_splitted",
26
    "src": [
27
      "$splitted"
28
    ],
29
    "action": "supervisely",
30
    "settings": {}
31
  }
32
]
Copied!
Here is the visualization of DTL query:
How to use neutral class
Find instances by segmenting object boundaries
Last modified 1yr ago
Copy link
Contents
Introduction
Task description
Step 1. Upload INRIA dataset
Step 2. DTL #1 to split data to train/val/test
Step 3. DTL #2 prepare training project
Step 3. UNet training
Step 4. Inference
Step 5. DTL #3 to split prediction mask to separate objects