Articles | Volume 12, issue 5
Research article
03 May 2018
Research article |  | 03 May 2018

Using machine learning for real-time estimates of snow water equivalent in the watersheds of Afghanistan

Edward H. Bair, Andre Abreu Calfa, Karl Rittger, and Jeff Dozier

Abstract. In the mountains, snowmelt often provides most of the runoff. Operational estimates use imagery from optical and passive microwave sensors, but each has its limitations. An accurate approach, which we validate in Afghanistan and the Sierra Nevada USA, reconstructs spatially distributed snow water equivalent (SWE) by calculating snowmelt backward from a remotely sensed date of disappearance. However, reconstructed SWE estimates are available only retrospectively; they do not provide a forecast. To estimate SWE throughout the snowmelt season, we consider physiographic and remotely sensed information as predictors and reconstructed SWE as the target. The period of analysis matches the AMSR-E radiometer's lifetime from 2003 to 2011, for the months of April through June. The spatial resolution of the predictions is 3.125 km, to match the resolution of a microwave brightness temperature product. Two machine learning techniques – bagged regression trees and feed-forward neural networks – produced similar mean results, with 0–14 % bias and 46–48 mm RMSE on average. Nash–Sutcliffe efficiencies averaged 0.68 for all years. Daily SWE climatology and fractional snow-covered area are the most important predictors. We conclude that these methods can accurately estimate SWE during the snow season in remote mountains, and thereby provide an independent estimate to forecast runoff and validate other methods to assess the snow resource.

Short summary
In Afghanistan, almost no snow measurements exist. Operational estimates use measurements from satellites, but all have limitations. We have developed a satellite-based technique called reconstruction that accurately estimates the snowpack retrospectively. To solve the problem of estimating today's snowpack, we used machine learning, trained on our reconstructed snow estimates, using predictors that are available today. Our results show low errors, demonstrating the utility of this approach.