DiGATe-UNet: Lightweight Dual-Stream Framework for Landslide Segmentation

November 5, 2025 · View on GitHub

This repository contains the implementation of our lightweight dual-stream Siamese framework for landslide segmentation from remote sensing imagery. The model integrates optical and topographical data fusion, an adaptive decoder with lightweight cross-attention, gated fusion, and deep supervision to achieve accurate boundary delineation and robust performance across diverse landscapes.

  • 📜 Our manuscript has been accepted by the CMC-Computers, Materials & Continua journal.

  • 📌 Current content: Full training pipeline, model implementation, and benchmark datasets integration.

Requirements

Tested on Python 3.10 with CUDA 12.x.

Minimal runtime dependencies (the ones actually used by the code/notebook):

numpy==1.26.4
torch==2.6.0
torchvision==0.21.0
timm==1.0.19
segmentation-models-pytorch==0.4.0
opencv-python==4.11.0.86
Pillow==11.1.0
h5py==3.13.0
tqdm==4.67.1
matplotlib==3.10.1
scikit-image==0.25.2

Optional (only if you use certain SMP backbones that rely on them):

efficientnet-pytorch==0.7.1
pretrainedmodels==0.7.4

Installation

  1. Create env (recommended)
conda create -n digate python=3.10 -y
conda activate digate
  1. Install PyTorch first (match your CUDA):
# Example for CUDA 12.x wheels
pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu121

If you’re on CPU or a different CUDA, use the command from the official PyTorch site for your setup.

  1. Install the rest
pip install -r requirements.txt

Quick Start (Inference) Weights required*

  1. Place the Bijie dataset (or your data) under the expected paths used in inference.ipynb.
  2. Open the notebook:
jupyter notebook inference.ipynb
  1. Run all cells to produce masks and basic metrics/visualizations.

Dataset

  • Bijie landslide dataset (examples used in the notebook).
  • Additional datasets and loaders will be wired in with the training pipeline update.

Project Structure (current)

└── \datset\                                  # Dataset preprocessing
└── \models\                                  # Models, Loss and metrics
└── \rrain\                                   # Training
└── \utils\                                   # Evaluate and plot
└── LS_segmentation_landslide4sense.ipynb     # Train on LandSlide4Sense dataset
├── inference.ipynb                           # Example inference workflow on Bijie
├── backbones.ipynb                           # Test computational efficieny on trained weigths
└── LS_segmentation_bijie.ipynb               # Train on Bijie dataset