MCD-Net: A Lightweight Deep Learning Baseline for Optical-Only Moraine Segmentation

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

Dataset: 10.5281/zenodo.18074779
Pre-trained Model: model_data/MCDNet_mobilenetv2_best.pth

---

Repository Structure

MCD-Net/
├── dataset/
│   └── Moraine_dataset/
│       ├── JPEGImages/               # Place your .jpg images here
│       ├── SegmentationClass/        # Place your .png masks here
│       └── ImageSets/Segmentation/   # Auto-generated txt splits
├── nets/                           
│   ├── attention.py                 
│   ├── deeplabv3_plus.py            
│   ├── deeplabv3_training.py       
│   └── mobilenetv2.py              
├── utils/                           
│   ├── callbacks.py                  
│   ├── dataloader.py             
│   ├── utils.py                   
│   ├── utils_fit.py               
│   └── utils_metrics.py          
├── model_data/                  
│   └── MCDNet_mobilenetv2_best.pth                                                                    
├── dataset_annotation.py           
├── train.py                       
├── mcdnet_predictor.py            
├── predict.py                      
├── get_miou.py                     
├── requirements.txt             
└── README.md                     

---

Installation

1. Clone the repository

git clone https://github.com/Lyra-alpha/MCD-Net.git
cd MCD-Net

2. Create environment & install dependencies

conda create -n mcdnet python=3.9 -y
conda activate mcdnet
pip install -r requirements.txt

3. Verify installation

python -c "import torch; print('PyTorch:', torch.__version__); print('CUDA available:', torch.cuda.is_available())"

---

Environment

Requirement	Version Tested
Python	3.9
PyTorch	1.12+
CUDA	11.6 (GPU recommended)
OS	Ubuntu 20.04 / Windows 10/11

GPU Memory: >= 6 GB VRAM (batch_size=8)

---

Dataset Preparation

1. Download the MCD Dataset from Zenodo.
2. Extract and place the files into:
   • dataset/Moraine_dataset/JPEGImages/ (input images, .jpg)
   • dataset/Moraine_dataset/SegmentationClass/ (label masks, .png)
3. Run the annotation script to generate the data splits:

   python dataset_annotation.py
   

   This creates train.txt, val.txt, test.txt, and all.txt inside dataset/Moraine_dataset/ImageSets/Segmentation/.

If you already have the split files, you can skip step 3 (but ensure they match the image filenames).

---

Training

Quick Start

python train.py

Key Configuration (in train.py)

Parameter	Default	Description
num_classes	2	Background + Moraine
backbone	mobilenet	Only MobileNetV2 supported
use_attention	True	Enable CBAM attention modules
downsample_factor	16	Output stride (8 or 16)
Freeze_Epoch / UnFreeze_Epoch	100 / 200	Two-stage training with SGD
Init_lr	7e-3	Initial learning rate
batch_size (Freeze)	8	Batch size during frozen backbone
batch_size (Unfreeze)	4	Batch size during full training

Training outputs are saved in logs/:

• best_epoch_weights.pth -- Best validation loss
• last_epoch_weights.pth -- Final epoch
• loss_*.png, epoch_miou.png -- Loss and mIoU curves

---

Prediction

Single Image Prediction

python predict.py
# Set mode = "predict" in predict.py
# Enter image path when prompted

Batch Folder Prediction (default)

python predict.py
# Set mode = "dir_predict" in predict.py
# Images from: dataset/Moraine_dataset/JPEGImages/
# Results saved to: img_out/

To use your own trained model, update model_path in mcdnet_predictor.py or pass it as a keyword argument.

---

Evaluation (mIoU, F1, Precision, Recall)

Evaluate on the validation set:

python get_miou.py

Results saved in miou_out/:

• mIoU.png, mPA.png, Recall.png, Precision.png, F1.png
• confusion_matrix.csv
• PixelAccuracy.png

---

Quick Test Example

# 1. Run data splitting
python dataset_annotation.py

# 2. Start training (2 epochs for sanity check)
#    In train.py, set: UnFreeze_Epoch = 2
python train.py

# 3. Evaluate
python get_miou.py

# 4. Predict on validation images
#    In predict.py, set:
#      mode = "dir_predict"
#      dir_origin_path = "dataset/Moraine_dataset/JPEGImages"
python predict.py

---

Pre-trained Model

File	Description
model_data/MCDNet_mobilenetv2_best.pth	Best checkpoint (provided)
model_data/mobilenet_v2.pth	ImageNet-pretrained MobileNetV2 (auto-download)

To use the pre-trained model for inference:

from mcdnet_predictor import MCDNetPredictor

predictor = MCDNetPredictor(
    model_path='model_data/MCDNet_mobilenetv2_best.pth',
    num_classes=2,
    use_attention=True
)

---

References

• DeepLabV3+ PyTorch Implementation
• CBAM: Convolutional Block Attention Module
• MobileNetV2: Inverted Residuals and Linear Bottlenecks

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

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