Monitoring earth’s glacial lakes from space with machine learning

The rapid worldwide formation and expansion of glacial lakes has increased the likelihood of glacial lake outburst floods, threatening lives and infrastructure, particularly in vulnerable mountain communities. Given the rapid increase in the popularity of artificial intelligence methods for remote sensing of glacial lakes, a comprehensive review is essential. We survey a decade (2015–2024) of research on glacial lake monitoring from space, with a focus on classical machine learning and deep learning approaches. We identify key trends, research gaps, and best practices for future studies. Most studies rely on optical imagery, especially Landsat-8 and Sentinel-2, while Sentinel-1 serves as a complementary radar source. However, monitoring glacial lakes in mountainous regions remains a challenge on cloudy days due to the limitations of radar and the unusability of optical data. Deep learning, particularly U-Net and DeepLab derivatives, dominates learning-based glacial lake studies but remains computationally demanding. Critical challenges involve balancing performance gains against trade-offs in data availability, computational cost, and model transferability. Geographic and methodological gaps, especially in regions experiencing rapid lake growth, underscore the need for broader spatial coverage and improved spatiotemporal model generalization. Moreover, transitioning from a focus on static seasonal mapping to frequent multi-temporal monitoring is beneficial for understanding glacial lake evolution and outburst flood hazards. Adapting emerging deep learning architectures to integrate multispectral, hyperspectral, and radar data could enhance glacial lake detection capabilities. Furthermore, thorough inter-method comparisons, benchmarking with rigorous evaluation metrics, and open-sourcing datasets and code would facilitate robust, large-scale glacial lake monitoring efforts. • Key regions with hazardous glacial lakes are understudied with learning-based methods. • U-Net and DeepLab dominate learning-based studies, but compute costs pose a barrier. • Emphasis on transferability assessments is needed to validate model generalization. • Fair benchmarking with strict evaluation metrics is crucial yet often overlooked.