DANN RTS Segmentation

This is the repo for out project Domain Adaptive Semantic Segmentation of Multi-annual Retrogressive Thaw Slumps (subject to change). Generally, we implemented a DeepLab V3+ segmentation model with Gradient Reversal Layer and a domain discriminator, in order to achieve transfer learning across remote sensing samples from multiple years.

Quick Start

For environment configuration, refer to file env.yaml

conda env create -f env.yaml

Note that there may be excessive requirements that are not needed and may slow down the configuration process. We may fix that later.

In a configured environment, one can simply run

python train.py

to begin training. For testing scripts, some related files are available upon request.

File Structure

The datasets folder should be placed outside this folder. During each time of training, our codes will create a folder outside this name logs, and create a sub-folder named with the time training starts, where weights, loss and accuracy plots are saved.

Pretrained Model

We adopted mobilenetv2 as the backbone. You can download it via https://pan.bnu.edu.cn/l/X1ns1x, and put it inside sub-folder model_data.

- datasets
-- 2019
--- images
--- labels
--- masks
-- 2020
-- ...

- DANN RTS Segmentation
-- model_data
--- deeplab_mobilenetv2.pth

- logs