The Use of Machine Learning to Identify Irrigated Locations and Estimate Applied Water Volume

A prime difficulty in ascertaining the impact of irrigation on the atmosphere is the paucity of irrigation data available. Regions that do report irrigation water volume tend to aggregate their data into seasonal or annual values, and many regions report none at all. To alleviate these issues, the authors propose a novel methodology leveraging satellite observed land surface temperature, numerical weather modeling, and machine learning to diagnose irrigation volume on daily time scales at spatial resolutions of 2-3 kilometers. This improves upon the spatial resolution offered by traditional water budget approaches using reanalysis products and satellite soil moisture observations (~25-50 km) and far surpasses the quarterly or annual irrigation reporting of official data sets. Preliminary results over Nebraska demonstrate that our machine learning approach has accuracy of 66% with a probability of irrigation detection of 74% compared to satellite-derived maps of irrigation. Inclusion of the diagnosed irrigation fields into the Weather Research and Forecasting model improves surface sensible heat flux RMSE by 17%-32% in irrigated regions and latent heat flux RMSE by up to 54%. Bias in land surface temperature decreases by up to 0.6 K in irrigated regions. Further, our approach naturally produces a range of possible irrigation volume, providing immediate insight into the uncertainty of the estimate.