Remote sensing of field-scale irrigation withdrawals in the central Ogallala aquifer region

Abstract

For agricultural areas facing water scarcity, sustainable water use policy relies on irrigation information that is timely and at a high resolution, but existing publicly available water use data are often insufficient for monitoring compliance or understanding the influence of policy on individual farmer decisions. This study attempts to fill this data gap by using remote sensing to map annual irrigation quantity at the field-scale within the central Ogallala aquifer region of the United States. We compiled in situ annual irrigation volume data at the field scale in the Republican River Basin of Colorado for 2015–2018 and at the Public Land Survey System (PLSS) section scale in western Kansas for 2000–2016, which served as reference data in random forest models that relied on Landsat-based actual evapotranspiration from the Operational Simplified Surface Energy Balance model (SSEBop) along with maps of irrigated area, Landsat spectral indices, climate, soils, and derived hydrologic variables. The models explained 87% of the variability in irrigation volume in Colorado and 75% in Kansas, but accuracy declined when transferring the models in spatial cross-validation (Colorado R2 =0.81; Kansas R2 =0.51) and temporal cross-validation (Colorado R2 =0.82; Kansas R2 =0.68). Predicted annual totals of irrigation volume in western Kansas had a mean absolute error of 11.9%, which was slightly higher than the average annual change of 11%. Use of predicted irrigation maps also lead to an underestimated effect size for a water use restriction policy in Kansas. These results indicate that field- and section-scale irrigation can be mapped with reasonable accuracy within a region and time period that has adequate sample data, but that methods may need to be improved for applying the models more broadly in areas that lack extensive in situ irrigation data to support further research on water use and aid in structuring policy.