Machine Learning Method Revolutionizes Supraglacial Lake Depth Measurement for Climate Research

As global warming accelerates, the increasing number of supraglacial lakes and the need to accurately measure their depths have become critical for understanding ice sheet mass balance and sea-level rise. These lakes, formed by meltwater accumulation on ice sheet surfaces, significantly influence ice sheet dynamics and melting rates. However, traditional methods for measuring their depths often struggle with accuracy, especially in deep lakes, and face challenges like the difficulty of in situ measurements and the limitations of satellite-based estimations.

A team from Sun Yat-sen University published a study in Journal of Remote Sensing, unveiling a novel method that combines machine learning with ICESat-2 satellite data and multispectral imagery. This new approach addresses the shortcomings of traditional techniques, providing a high-precision solution for large-scale monitoring of supraglacial lake depths. By integrating machine learning algorithms with advanced satellite data, this method overcomes the challenges faced by conventional models, offering a more accurate and scalable way to estimate lake depths.

The study introduces a unique approach that combines machine learning algorithms, such as XGBoost and LightGBM, with ICESat-2 satellite data and multispectral imagery from Landsat-8 and Sentinel-2. The core innovation lies in the application of these advanced algorithms, which significantly improve depth estimation accuracy. For example, XGBoost achieved a root mean square error of just 0.54 meters when applied to Sentinel-2 L1C imagery, marking a substantial improvement over traditional methods like the radiative transfer equation and spectral band ratio, which often struggle with accuracy, especially in deeper lakes.

The research team developed an enhanced Automated Lake Depth algorithm to extract reliable lake depth sample points from ICESat-2 ATL03 data. These points were then paired with multispectral imagery to generate training data for machine learning models. Testing the method on seven supraglacial lakes in Greenland, the results showed that the machine learning algorithms, particularly when using Sentinel-2 L1C imagery, offered the most accurate depth estimates. The study also explored the impact of atmospheric corrections on depth retrieval, finding that top-of-atmosphere reflectance data performed better than atmospherically corrected data for mapping lake bathymetry.

By employing an improved ALD algorithm to process ICESat-2 ATL03 data and pairing it with multispectral imagery from Landsat-8 and Sentinel-2, the team created a powerful tool for glacier lake monitoring. This methodology was tested against reference depths from the ALD algorithm and validated using ArcticDEM data. The high accuracy and scalability of the machine learning-based approach open up new possibilities for large-scale monitoring in polar regions and other glaciated areas, crucial for assessing climate change impacts. The full study is available at https://spj.science.org/doi/10.34133/remotesensing.0416.