A Deep Learning-based Toolbox for Automated Monitoring of Central Asian Glacial Lakes from Space

In recent decades, climate change has intensified the melting of glaciers in high mountain regions around the world, leading to the formation of new glacial lakes. These lakes can cause damage up to several hundred kilometres downstream when an outburst flood occurs. While more and more glacier lake inventories are becoming available to the research community, high-frequency mapping and monitoring of these lakes are still essential to assess hazards and estimate Glacial Lake Outburst Flood (GLOF) risks, particularly for lakes with high seasonal variations. In Central Asia, new lakes have been known to develop quickly, and non-stationary lakes can expand or regrow within a matter of weeks to months. Monitoring these lakes is crucial to understanding and mitigating the risks they pose.

Detecting glacial lakes using satellite sensors is difficult due to their small size and the fact that they are often frozen for much of the year. Furthermore, optical satellite imagery can be hindered by clouds. Additionally, cast and cloud shadows, as well as increasing lake and atmospheric turbidity, make it challenging to accurately observe and monitor these lakes using optical satellite imagery. On the other hand, using a SAR satellite sensor to monitor these lakes is difficult during windy scenarios and changes in backscattering due to variations in turbidity and the presence of cast shadows.

We have developed a Python-based toolbox for mapping potentially dangerous glacial lakes in Central Asia and for monitoring the dynamics of these lakes over time and space. The proposed analytical toolbox uses a combination of Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical satellite data as input. Satellite data fusion allows high temporal resolution, while SAR can penetrate cloud cover and allow year-round monitoring. The user interface for the toolbox is designed to also accommodate users with a non-programming background.

The Convolutional Neural Network (CNN)-based approach fuses information from heterogeneous satellite input data by learning joint satellite embeddings (feature representations), that are equivariant to the type of satellite input data. The proposed network has separate encoder branches for each input sensor. The learned embeddings are then fused to guide the identification of glacial lakes. The ultimate goal of our data-driven methodology is to create geolocated maps of the target regions by classifying each pixel as either a lake or background in a supervised manner.

This work is part of the GLOFCA project which aims to lower the vulnerability of people in Kazakhstan, Kyrgyzstan, Tajikistan, and Uzbekistan to GLOFs. This project is implemented by UNESCO and funded by the UN Adaptation Fund, in collaboration with various international and national partners.