Speaker: Yifan Sun
Affiliation: EPSS, UCLA
Date: Wednesday, June 3, 2026
Time: 12:00 PM
Abstract
Quantifying spatiotemporal dynamics of landslides, including occurrence, reactivation, and timing, is critical for understanding hazard risk in steep mountains. However, in regions like the Eastern Himalayas, mapping these spatiotemporal patterns is heavily restricted by persistent cloud cover, land-cover variability, and a scarcity of historical ground-truth inventories. To overcome these data and observational limitations, we utilized a novel unsupervised machine learning framework—the Stable Diffusion Change Detector (SDCD)—to extract continuous surface disturbance records directly from time-series optical satellite imagery. By bypassing the limitations of traditional manual labeling, this approach allows us to generate high-resolution landslide inventories in a data-scarce, cloud-prone region. Applying this method, we identify ~300 surface disturbance events, including landslides in hillslopes and channel-adjacent regions, over 1300 km2 and 10 years. We investigate temporal triggering of these events by analyzing daily precipitation and spatial patterns by comparing environmental conditions, including precipitation, long-term erosion rates, and topographic slope. Our time-series analyses indicate that abundant landslides in frontal basins are directly related to the magnitude of daily precipitation events, and the spatial density of overall landslides is highly consistent with independently measured long-term erosion rates. These results highlight the importance of high-resolution landslide datasets, both spatially and temporally, for understanding the complex controls on landscape change over time.