Remote Sensing Sar-Optical Land-use Classfication Pytorch
Refer to https://github.com/AmberHen/WHU-OPT-SAR-dataset.
Datasets：Sar and Optical
https://pan.baidu.com/s/1sIGsD3lBEogSCqzbDOaclA password：i51o
Paper Link: MCANet: A joint semantic segmentation framework of optical and SAR images for land use classification
https://www.sciencedirect.com/science/article/pii/S0303243421003457
- Release Crop Code (sar, opt, lbl)
- Release Convert Label Code
- Release Split Code (6:2:2) -> (17640:5880:5880)
- Upload to Server (four GPUs)
- The project should be organized as:
SOLC
├── dataset
| |whu-opt-sar-dataset-256 // root
│ ├── ├── train
│ ├── │ ├── sar
│ ├── │ │ ├── NH49E001013_1.tif
│ ├── │ ├── opt
│ ├── │ │ ├── NH49E001013_1.tif
│ ├── │ ├── lbl
│ ├── │ │ ├── NH49E001013_1.tif
│ ├── ├── val
│ ├── │ ├── sar
│ ├── │ │ ├── NH49E001013_2.tif
│ ├── │ ├── opt
│ ├── │ │ ├── NH49E001013_2.tif
│ ├── │ ├── lbl
│ ├── │ │ ├── NH49E001013_2.tif
│ ├── ├── test
│ ├── │ ├── sar
│ ├── │ │ ├── NH49E001013_3.tif
│ ├── │ ├── opt
│ ├── │ │ ├── NH49E001013_3.tif
│ ├── │ ├── lbl
│ ├── │ │ ├── NH49E001013_3.tif
├── libs // utils
├── models // model
├── tools // preprocessing
├── dataset.py
├── class_names.py
├── palette.py
├── sync_transforms.py
├── train.py
├── _test.py / predict.py
- Release Deeplabv3+ Code (sar + opt, 5-channels input)
- Release Learning Strategy Code (step size, gamma)
- Release Training Torch-Parser Code
- Release VGG19 Code (based on official implement) weights
- Release MCANet Code (unofficial implement, based on deeplabv3+)
- Release FCNs Code (FCN8s, 16s and 32s)
- Release Resnet-50 weights , Resnet-101 weights, and Restnet-152 weights Code
- Release Unet Code
- Release Segnet Code
- Release PSPnet Code
- Release SOLC Code (Ours)
- Release SOLC V1 Code (based on RGB-D and dual-resnet 50)
- Release SOLC V2 Code (based on dual-stream deeplabv3+)
- Release Deeplabv3+ Performance (Training 7 hours)
- Release Train Code
- Release Predict/Test Code
- Release SOLC V1 Code Performance
- Release SOLC V2 Code Performance
- Release SOLC V3 Code (based on dual-stream deeplabv3+ and SAGate)
- Release SOLC V4 Code (based on dual2one-stream deeplabv3+)
- Release SOLC V5 Code (based on dual2one-stream deeplabv3+, SAGate and RFB)
- Release SOLC V6 Code (based on dual2one-stream deeplabv3+ and Two Enhanced Module)
- Release SOLC V7 Code (based on dual-stream deeplabv3+, SAGate and ERFB) (successful version, Congratulations!)
- Retrain Unet, Segnet, and MCANet
- Retest the performance
- Release Our SOLC V7 weights solcv7: baiduyun password：solc
- Release Other model weights others: baiduyun password：solc
- Release 像素统计
[0, '背景', 11824154.0, 1.0] (忽略)
[1, '农田', 708762053.0, 0.016682825992096393]
[2, '城市', 96237141.0, 0.12286476797975535]
[3, '村庄', 119738993.0, 0.09874940237721894]
[4, '水体', 292127191.0, 0.04047604729817842]
[5, '森林', 774385740.0, 0.015269075073618998]
[6, '道路', 19661970.0, 0.6013717852280317]
[7, '其他', 35164414.0, 0.3362534066400197]
weight in criterion: [1.0, 0.016682825992096393, 0.12286476797975535, 0.09874940237721894, 0.04047604729817842, 0.015269075073618998, 0.6013717852280317, 0.3362534066400197]- Release More Results
- Release Plot Code
- 设置合适的空洞卷积膨胀率atrous_rates
- 余弦退火重启动学习率策略warm up
- 使用更多的数据增强
- 使用更强的损失函数（focal loss）或者为类别赋予权重（见tools/class_weight.py） 来解决类别不平衡问题
| 策略 | 模型 | 总体性能 |
|---|---|---|
| epoch=40, batch size=16, Random Flip, lr=1e-3, wd=1e-4 |
deeplabv3+ (pretrained=False) | oa=0.8096,mIoU=0.4118,kappa=0.7261 |
| epoch=40, batch size=16, Random Flip, lr=1e-3, wd=1e-4 |
unet (pretrained=False) | oa=0.7846,mIoU=0.3847,kappa=0.6826 |
| epoch=40, batch size=16, Random Flip, lr=1e-3, wd=1e-4 |
segnet (pretrained=False) | oa=0.7370,mIoU=0.3139,kappa=0.6148 |
| epoch=40, batch size=16, Random Flip, lr=1e-3, wd=1e-4 |
mcanet (pretrained=False) | oa=0.8179,mIoU=0.4296,kappa=0.7352 |
| epoch=40, batch size=16, Random Flip, lr=1e-3, wd=1e-4 |
solcv7 (pretrained=False) | oa=0.8222,mIoU=0.4382,kappa=0.7428 |
| 模型 | farmland | city | village | water | forest | road | others | background(ignored) |
|---|---|---|---|---|---|---|---|---|
| deeplabv3+ | 0.7951/0.6314 | 0.9414/0.6578 | 0.3928/0.3291 | 0.7856/0.5511 | 0.9422/0.8843 | 0.3504/0.2231 | 0.0269/0.0173 | 0.0/0.0 |
| unet | 0.7970/0.6081 | 0.5671/0.5174 | 0.3739/0.2755 | 0.7517/0.5758 | 0.9579/0.8418 | 0.5186/0.2467 | 0.0216/0.0120 | 0.0/0.0 |
| segnet | 0.7248/0.6122 | 0.0083/0.0083 | 0.4314/0.2660 | 0.8837/0.4845 | 0.9691/0.8206 | 0.3984/0.2948 | 0.14046/0.0249 | 0.0/0.0 |
| mcanet | 0.8320/0.6499 | 0.8394/0.6575 | 0.3892/0.3318 | 0.8177/0.6243 | 0.9476/0.8724 | 0.5036/0.2865 | 0.0169/0.0145 | 0.0/0.0 |
| dcn(ours) | 0.8392/0.6505 | 0.9051/0.6880 | 0.3999/0.3337 | 0.7979/0.6035 | 0.9425/0.8878 | 0.5251/0.3250 | 0.0192/0.0167 | 0.0/0.0 |
| 模型 | OA | mIOU | kappa |
|---|---|---|---|
| deeplabv3 | 0.8096 | 0.4118 | 0.7261 |
| unet | 0.7846 | 0.3847 | 0.6826 |
| segnet | 0.7370 | 0.3139 | 0.6148 |
| mcanet | 0.8179 | 0.4296 | 0.7352 |
| dcn(ours) | 0.8222 | 0.4382 | 0.7428 |
左侧为各类别的平均精度OA，右侧为各类别的平均mIouNote: 0->background，10->farmland，20->city，30->village，40->water，50->forest，60->road，70->others- Clone this repo.
$ git clone https://github.com/yisun98/SOLC.git
$ cd SOLC- Install Environments
$ pip install -r requirements.txt
$ source activate- Dataset
$ python tools/crop_sar.py
$ python tools/crop_opt.py
$ python tools/convert_lbl.py
$ python tools/crop_lbl.py
$ python tools/split_data.py- Training
nohup python train.py >> train_<model_name>.out 2>&1 &Please see train.py for details.
tensorboard --logdir=<your_log_dir> --bind_all - Test/Predict
nohup python train.py --model solcv7 --num_classes 8 >> train_<model_name>.out 2>&1 &Please see train.py for details.
python predict.py --model solcv7 --model-path <model_path>More results please see here password：solc.
If you have any questions, welcome to contact us !
Contributors 1: yisun98 - [email protected]
Contributors 2: yiruzzz - [email protected]
Contributors 3: Bismarckwzc - [email protected]
Contributors 4: 917596622 - [email protected]
If the project is helpful to you, please consider citing us.
@misc{SOLC2022,
author = {Y. Sun, Y. Zhao, Z. Wang, Y. Fan},
title = {SOLC},
year = {2022},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/yisun98/SOLC}},
}
