Official implementation of MCD-Net , a lightweight deep learning framework for optical-only moraine segmentation, as presented in our IEEE journal paper.

MCD-Net is a lightweight deep learning baseline that integrates MobileNetV2, Convolutional Block Attention Module (CBAM), and DeepLabV3+ decoder for moraine segmentation from optical imagery. This work establishes the first reproducible benchmark for optical-only moraine segmentation with a novel dataset of 3,340 annotated high-resolution images.

The MCD Dataset contains 3,340 high-resolution image-mask pairs from Sichuan and Yunnan, China:

• Images: 1024×1024 pixels, 0.5-2.0m resolution
• Classes: Binary segmentation (background vs. moraine body)
• Split: 2,630 training + 293 test images

Download the dataset from: https://doi.org/10.5281/zenodo.18074779

1. Place the dataset downloaded from Zenodo into the dataset folder.
2. Before training, place the label files in dataset/Morainse_dataset/SegmentationClass and the image files in dataset/Morainse_dataset/JPEGImages.
3. Run dataset_annotation.py to generate the corresponding dataset split text files before training.
4. In the train.py file, select the pre-trained weights you want to use (default parameters are already set).
5. Run train.py to start training.

This repository provides a trained pth file (MCDNet_mobilenetv2_best.pth). Set the relevant paths in mcdnet_predictor.py, then select the prediction mode in predict.py and run it. If you want to use your own trained model, please modify the relevant paths accordingly.

https://github.com/ggyyzm/pytorch_segmentation

https://github.com/bubbliiiing/deeplabv3-plus-pytorch

https://github.com/bonlime/keras-deeplab-v3-plus