Image Classification

This is classification turtorial for satellite image. We will use sentinal-2 TCI data as an example. It cover from training data prepare, train the model, and predict the new files.

Prepare trainning data

Suppose we got a scene of sentinel-2 satellite TCI image data. You can download from esa scihub. I has got the secene id T51RTQ_20200513T023551_TCI and convert the JPEG2000 to GeoTIFF.

  1. patch the large 10980x10980 pixel image to 128x128 pixel image

    ➜ cd  tests/classification/      
    ➜ ts make-mask-cls --filepath T51RTQ_20200513T023551_TCI.tif --width 128 --height 128 --outpath ./patched
    processing 1/1 file T51RTQ_20200513T023551_TCI.tif ...
    14%|███████████████▉                                                              | 12/85 [00:07<00:45,  1.60it/s]

You should get the following data:

../_images/classification_patch.png

  1. labeling the train data and test data

You can split the data into tran data and test data as below. And then labeling those patched image into four classes: water, residential, farmland, forest. Reorganize the catalog of these small images according to different categories and split them to train and validation dataset.

    .
    ├── train
    │    ├── water
    │    │   ├── T51RTQ_20200513T023551_TCI_1_29.tif
    │    │   ├── T51RTQ_20200513T023551_TCI_2_29.tif
    │    │   └── ...
    │    ├── frameland
    │    │   ├── T51RTQ_20200513T023551_TCI_3_2.tif
    │    │   ├── T51RTQ_20200513T023551_TCI_11_1.tif
    │    │   └── ...
    │    ├── residential
    │    │   ├── T51RTQ_20200513T023551_TCI_0_29.tif
    │    │   ├── T51RTQ_20200513T023551_TCI_1_37.tif
    │    │   └── ...
    │    ├── forest
    │    │   ├── T51RTQ_20200513T023551_TCI_7_21.tif
    │    │   ├── T51RTQ_20200513T023551_TCI_22_45.tif
    │    │   └── ...
    ├── validation
    │    ├── water
    │    │   ├── T51RTQ_20200513T023551_TCI_5_32.tif
    │    │   ├── T51RTQ_20200513T023551_TCI_5_12.tif
    │    │   └── ...
    │    ├── frameland
    │    │   ├── T51RTQ_20200513T023551_TCI_9_2.tif
    │    │   ├── T51RTQ_20200513T023551_TCI_6_76.tif
    │    │   └── ...
    │    ├── residential
    │    │   ├── T51RTQ_20200513T023551_TCI_8_29.tif
    │    │   ├── T51RTQ_20200513T023551_TCI_3_37.tif
    │    │   └── ...
    │    ├── forest
    │    │   ├── T51RTQ_20200513T023551_TCI_8_12.tif
    │    │   ├── T51RTQ_20200513T023551_TCI_31_5.tif
    │    │   └── ...

Training model

You can use the scipts/train_cls.py to train your classification model. Here is a train examples, for more parameter information, please use python3 train_cls.py help to view.

python3 train_cls.py --train-path ./classification/train \
                 --val-path ./tests/classification/val \
                 --model resnet18 \
                 --batch-size 16 \
                 --num-classes 4 \
                 --device cuda

It should start train and the console will print the training logs.

Train Epoch: 0 [0/1514 (0%)]    Loss: 1.878198
Train Epoch: 0 [160/1514 (11%)] Loss: 0.811605
Train Epoch: 0 [320/1514 (21%)] Loss: 0.774963
Train Epoch: 0 [480/1514 (32%)] Loss: 0.817051
Train Epoch: 0 [640/1514 (42%)] Loss: 0.869388
Train Epoch: 0 [800/1514 (53%)] Loss: 4.763704
Train Epoch: 0 [960/1514 (63%)] Loss: 0.968885
Train Epoch: 0 [1120/1514 (74%)] Loss: 4.856205
Train Epoch: 0 [1280/1514 (84%)] Loss: 1.343379
Train Epoch: 0 [1440/1514 (95%)] Loss: 0.551179

Test set: Average loss: 16.4018, Accuracy: 68/326 (21%)

Train Epoch: 1 [0/1514 (0%)]    Loss: 2.768502
Train Epoch: 1 [160/1514 (11%)] Loss: 0.424574
Train Epoch: 1 [320/1514 (21%)] Loss: 0.572497
Train Epoch: 1 [480/1514 (32%)] Loss: 1.539173
Train Epoch: 1 [640/1514 (42%)] Loss: 0.707925
Train Epoch: 1 [800/1514 (53%)] Loss: 0.545577
Train Epoch: 1 [960/1514 (63%)] Loss: 0.956915
Train Epoch: 1 [1120/1514 (74%)] Loss: 0.556552
Train Epoch: 1 [1280/1514 (84%)] Loss: 0.825140
Train Epoch: 1 [1440/1514 (95%)] Loss: 0.588212

Test set: Average loss: 0.4656, Accuracy: 254/326 (78%)

Train Epoch: 2 [0/1514 (0%)]    Loss: 0.422114
Train Epoch: 2 [160/1514 (11%)] Loss: 0.273431
Train Epoch: 2 [320/1514 (21%)] Loss: 0.505005
....

You can also open Tensorboard to view the loss and test accuracy during training. ../_images/tensorboard-classification.png

And we provide lots of classification model, and all of them support multi-channel(>4) tiff image as input.

  • VGG: vgg11, vgg11_bn, vgg13, vgg13_bn, vgg16, vgg16_bn, vgg19_bn, vgg19
  • ResNet: resnet18, resnet34, resnet50, resnet101, resnet152, resnext50_32x4d,'resnext101_32x8d, wide_resnet50_2, wide_resnet101_2
  • DenseNet: densenet121, densenet169, densenet201
  • Inception: inception_v3
  • MobileNet: mobilenet_v2
  • EfficientNet: efficientnet_b0, efficientnet_b1, efficientnet_b2, efficientnet_b3,'efficientnet_b4, efficientnet_b5, efficientnet_b6, efficientnet_b7
  • ResNeSt: resnest50, resnest101, resnest200, resnest269

Inference new image

[TODO]