Dirty road extraction from GF-2 images by semi-supervised deep learning method for arid and semiarid regions of southern Mongolia The experimental setup is forked from UniMatch(https://github.com/LiheYoung/UniMatch).
Due to the limitation of upload data and model size, we store the training data set and the trained model in Baidu online disk, which is linked as https://pan.baidu.com/s/1vu2lD-qfuXxYij8r-MOSSg (Extraction code:u9rc). For training, please directly extract the three files from the zip file into your home directory. Then follow the training instruction in this instruction file.
Comparison utilizing exclusively labeled data.
| Method | MeanIou | back_iou | road_iou |
|---|---|---|---|
| Xception(sup) | 84.1 | 99.12 | 69.07 |
| Resnet101(sup) | 85.48 | 99.19 | 71.78 |
| Resnet101(unimatch) | 86.21 | 99.2 | 73.22 |
Comparison with various unlabeled data.
| Method | MeanIou | back_iou | road_iou |
|---|---|---|---|
| Resnet101(unimatch) | 86.21 | 99.2 | 73.22 |
| Resnet101(unimatch) | 86.37 | 99.24 | 73.51 |
The checkpoints are situated within the 'experiments' folder.
cd UniMatch
conda create -n unimatch python=3.10.4
conda activate unimatch
pip install -r requirements.txt
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 -f https://download.pytorch.org/whl/torch_stable.html├── ./pretrained
├── resnet101.pth
└── xception.pth
├── data
├── roadseg_semi_new2
└── train
└── val
└── txts
bash scripts/train_new2_uni_res_1k.sh 4 12360bash scripts/train_new2_sup_res_b8.sh 4 12360
bash scripts/train_new2_sup_xcep_b8.sh 4 12360