Machine learning on images: Combining passive microwave and optical data to estimate snow water equivalent in Afghanistan’s Hindu Kush

Feb. 25, 2016 from 3:30 to 4:20 p.m. — More Hall, room 220

Jeff Dozier

Professor, School of Environmental Science & Management, UC Santa Barbara


The task is to estimate the spatiotemporal distribution snow water equivalent (SWE), and thereby spring snowmelt runoff, in snow-dominated mountain environments, including those that lack on-the-ground measurements such as the Hindu Kush range in Afghanistan. During the snow season, two remotely sensed datasets are available: (1) passive microwave calculations of SWE generally underestimate in the mountains; (2) snow-covered area measurements are accurate but do not provide SWE. Another dataset is available after the snow has melted. From the date the snow disappears, we can reconstruct the accumulated SWE back to the last significant snowfall by calculating the energy used in melt. This reconstructed SWE provides a training set for predictions from the passive microwave SWE and snow-covered area, so we can use machine learning methods to discover patterns to try to estimate SWE from the data that are available during the snow season. Several methods work—regression-boosted decision trees, bagged trees, neural networks, and genetic programming—with R2 values around 0.8. Predictors built with multiple years of data reduce the bias that usually appears if we predict one year from just one other year’s training set. Adding other precipitation estimates, for example from the Global Precipitation Measurements mission, is in progress.


Professor in the Bren School of Environmental Science & Management at UC Santa Barbara.