Glacial-Lakes

In this repository, you will find code and resources related to lake detection using deep learning techniques. The project leverages satellite imagery and advanced neural network architectures to accurately identify and segment lakes in various geographical regions.

Project Overview

Accelerated glacier retreat driven by global climate change has led to the rapid formation and expansion of glacial lakes in alpine environments, significantly increasing the risk of Glacial Lake Outburst Floods (GLOFs). This research proposes a two-part methodological framework to improve the detection and monitoring of these dynamic water bodies using multi-source remote sensing data.

First, a Melting Season Assessment was developed using Sentinel-1 Synthetic Aperture Radar (SAR) backscatter data to automatically identify the onset of the melting season. This period is critical for hazard assessment as it marks the time when newly formed or rapidly expanding lakes pose the highest risks. The results indicate that the melting season start can be reliably detected using $\sigma^{0}_{VV}$ backscatter with a 3 dB threshold and a 25% wet snow percentage threshold. This method achieved high accuracy, identifying the start of snowmelt within an average of 6.5 days compared to in-situ physical simulations.

Results Example:

Second, Glacial Lake Segmentation was performed using deep learning-based semantic segmentation models. These models integrated multi-source data, including Sentinel-1 (radar), Sentinel-2 (optical), and Digital Elevation Models (topographical). The study evaluated several architectures, including U-Net variants, DeepLabV3+, and a hybrid SwinV2-CNN U-Net. The findings demonstrate that the hybrid SwinV2-CNN U-Net was the best-performing model, effectively capturing both global context and fine spatial details necessary for accurate segmentation of early-stage lakes.

Overall, the work demonstrates that advanced deep learning models outperform traditional machine learning methods for glacial lake mapping.

The best performing model (SwinV2-CNN U-Net) architecture:

Repository Structure

• dataset/: Contains the data for the Deep Learning models, including satellite images and corresponding labels.
• code/: Houses all the scripts and notebooks used for data preprocessing, model training, evaluation, and visualization.
  • melting_season/: Notebooks and scripts specific to the melting season analysis.
  • deep_learning_dataset/: Scripts for handling and reconstructing the images dataset from GEE.
  • lake_detection/: Code related to lake detection models and experiments used as a Classification task.
  • lake_detection_deep_learning/: Contains deep learning models for lake detection, including model architectures, training routines, and visualization tools.
    • models/: Pre-trained models and checkpoints for various experiments.
    • trainer/: Is the library for training inside of it are all the models architectures.

Author

Tim Ernst

Scope of the Project

This work is part of the Master's thesis "Detection of glacial lakes using remote sensing data" for the degree of Master of Science in Data Science at the University of Applied Sciences and Arts Western Switzerland (HES-SO).