Multi-Sensor Glacier Surface Classification Using Confidence-Aware Explainable Inverse-Mapping Neural Network

Abstract

Mapping snow cover at the end of the ablation season allows us to extract the snow line altitude (SLA). The SLA is an important proxy for the equilibrium line altitude of a glacier and an indicator of glacier health. With the increase in both active and passive remote sensing satellites, the accuracy and effectiveness of glacier monitoring can be enhanced, as the two sensors offer complementary information. In this paper, we focus on the combination of Sentinel-1 synthetic aperture radar (SAR) and Sentinel2 optical data to perform glacial classification using an explainable neural network and thereafter determine SLA. Additionally, confidence-aware inverse mapping dynamics is used to understand the result reliability and the individual sensor contributions. The proposed method is applied to the Great Aletsch Glacier in the European Alps, where an overall accuracy of 83% is observed compared to the ground truth data. We observe the glacier from 2015 to 2023, noting a retreat of the SLA to higher elevations by 36 to 133 m depending on the region. Apart from climaterelated mass loss, the European Alps are also affected by dust deposited during Sahara dust events (SDE) and contamination from algae. Thus, in this work, we assess the annual presence of contaminated snow on the glacier. The inverse mapping dynamics reveals the contributions of both SAR and optical sensor data in the classification. This multi-sensor approach is shown to mitigate the limitations of single-source data, providing a comprehensive understanding of glacier dynamics in the context of climate change.