Abstract
Climate warming is a global phenomenon with particularly pronounced impacts in high-altitude mountainous regions. These areas are experiencing rapid transformations, with glacial and permafrost retreat producing increasingly evident direct effects on the landscape. One significant outcome of these changes is the growing frequency of landslides in glacial environments, which reflects a paraglacial response to ice loss and permafrost degradation. The accumulation of debris on glaciers alters their thermal regime, impacts debris distribution, and affects alpinist routes also endangering high altitude infrastructures like huts and bivouacs. Identifying the location and frequency of these events is thus a critical step for assessing areas vulnerable to destabilization and understanding their connection to external triggers. Here, we propose a tool implemented in Google Earth Engine and Python that combines spaceborne radar and multispectral data to detect and classify glacial changes, minimizing the limitations inherent in each sensor type. Our workflow, tested on the glacial surfaces of Vedretta della Miniera (Val Zebrù, Italy), Tscherva Glacier (Val Roseg, Switzerland), and Mount Cook (New Zealand) in different snow conditions, analyzes changes in the Normalized Difference Snow Index (NDSI) derived from Sentinel-2 (S2) within specific time ranges to extract preliminary debris maps. It then integrates Sentinel-1 (S1) backscatter information to fill information gaps caused by cloud coverage and provides refined information of surface changes. The methodology consists of two main analytical blocks. First, we filter Sentinel-2 optical images in a user defined time range and over the glacial surface (Randolf Glacier Inventory extent; RGI Inventory, 2023) inside the selected area of interest. We then exclude cloudy pixels and calculate the normalized difference snow index (NDSI) for all the others applying a threshold to differentiate between areas likely covered by snow or ice (NDSI>0.3) and those corresponding to bare rock (NDSI<0.3). To improve rock pixel classification accuracy and avoid misidentifying temporarily covered rocks such as nunataks and bedrock outcrops, we compare each pixel's NDSI value with its value during the closest maximum ablation period. This comparison excludes pixels that already show rock signatures when snow cover is at its minimum. The union of pixels thus identified generates an initial debris extension map, with uncertainties stemming from the individual steps of cloud detection, snow and rock recognition. According to cloud extent in the consider time span, especially during winter season, a lot of pixels can remain not classified and thus prevent a correct detection of surface changes. We thus consider Sentinel-1 GRD intensity images (Mulissa et al., 2021) selected within the same user-defined timespan to compute the VH backscattering backward difference (∆dB) over all those areas that have not been classifies as debris from S2 analysis. Using a size-dependent filter, tailored to the minimum landslide size we aim to detect, we outline discrete pixel clusters and compute the mean ∆VH within each of them. Clusters whose mean ∆VH falls outside the interquartile range (IQR3) of the cluster values in the image indicate the most prominent changes on the glacier. By comparing this value with the VH value from the previous year’s accumulation period and applying a classification method based on a Support Vector Machine (SVM) model, we can distinguish between ice and snow cover. This allows us to isolate pixels that are more likely associated with debris accumulation, rather than changes in snow or ice. Integrating S1 and S2 data enables the creation of a comprehensive map of new debris accumulation, minimizing uncertainties and reducing false positives. Our results demonstrate a strong correlation between the identified clusters and manually mapped landslide deposits, which were used as the reference extent. For instance, we observed a 90% overlap between the landslide cluster in Vedretta della Miniera and Mt. Cook, where the debris deposits were manually mapped shortly after the event. In contrast, the overlap with Mt. Scerscen was 50%, where manual mapping from satellite images was only possible two months post-event due to persistent snow and cloud coverage. This delay led to the remobilization and rearrangement of the initial debris deposit, affecting the accuracy of the overlap. The tool leverages open-source libraries and datasets, making it readily adaptable to other glacial environments.