Knowledge Bank

The Cordillera Blanca, Peru contains 25% of the world's tropical glaciers and simultaneously acts as a natural water tower for the surrounding population centers and one of the largest sources of the Amazon River. Prior studies and expeditions have shown that this region is in a state of rapid change as the glaciers withdraw at an accelerating pace. However, there is a lack of accurate glacier and lake map inventories for the region, something that is vital for a rapidly evolving landscape that is prone to hazards and needs to adapt to altered water reservoirs. This issue is, in part, driven by the difficult task of manually digitizing glacier and lake outlines. This research project leverages four decades of NASA Earth observation data alongside recent advances in cloud computing and machine learning to overcome remote sensing challenges to automate the creation of these outlines. I first subject satellite imagery collected over the Cordillera Blanca since 1987 to temporal, spatial, and cloud-filtering algorithms. Surviving images are then used to calculate annual median composite images. Indices are then calculated that help to distinguish water phases and debris cover from the surrounding terrain when applied to composites that include synthetic aperture radar data where possible. The Random Forest classifier uses these to generate an inventory of annual glacier and lake outlines that are then validated via model evaluation, comparison to previous inventories, and a novel pixel-summing confidence technique. This research represents a significant step toward generating accurate, cloud-free satellite imagery time-series data products on-demand via cloud computing and Random Forest algorithms.