Classifying Irrigation: Developing a CONUS-wide Neural Network Classifier

Abstract:

Irrigation is the largest consumptive user of freshwater on the planet, often unsustainably depleting groundwater resources Stakeholders and policymakers in locations where the impacts of unsustainable water are felt most tangibly, such as California’s Central Valley, have begun implementing various conservation strategies. These strategies leverage scientific input for optimizing water usage in order to maintain irrigated agricultural production. However, even with this pressure to better manage regional aquifer resources, a cohesive and contiguous United States Earth observation-derived high spatiotemporal resolution irrigation product remains elusive. In particular, many irrigation products experience decreased accuracy within humid regions (e.g. the Midwest), others struggle with classifying relatively widely-spaced orchard crops. Here we test a cloud computing based framework for annually classifying irrigation across the contiguous United States (CONUS), at high-resolution (30 meter), over multiple decades. This framework deploys an artificial neural network, trained on a broad and diverse database of observations, driven with expertly-selected inputs derived from Earth observations and climatological reanalysis products within Google Earth Engine (GEE). The selected inputs contain recently-developed spectral indices shown to detect nuanced differences between irrigated and non-irrigated fields in previous work. After training and running the classifier, we compare the outputs to other currently available products to evaluate strengths and differences between classification approaches. This framework lays the groundwork for understanding ongoing changes in agricultural landuse and provide valuable input for data-driven agrohydrological models that can be used to inform sustainable land management practices.