Classifying Irrigation Using Cloud Computing: An Overview of an Earth Observation powered Neural Network Classifier

Abstract:

Accelerating human population growth, rising standards of living, and increasingly more interwoven international agricultural trade patterns generate tremendous pressure toward global food security. A predominant technological solution to these pressures historically has been the widespread adoption of irrigation. As a consequence, irrigation has become the largest global consumer of freshwater; in many heavily-irrigated aquifers this consumption occurs at unsustainable levels. Various approaches have been used to quantify the timing and complex evolution of irrigated agricultural lands in the United States. Currently-available data products have been generated from coarse-resolution reported data (county and state level) derived from the 5 year recurrent USDA NASS Agricultural census and input from a number of Earth observation systems (e.g. Landsat, Sentinel, MODIS). The accuracies, extent, and spatial resolution of these data products vary, which affects their reliability for discerning nuanced changes through time, their robustness in driving process-based hydrologic models, and their useability for informing the decision making of stakeholders and policymakers. To address these limitations, we developed a framework to generate a contiguous United States (CONUS), high-resolution (30 meter), multidecadal dataset by utilizing the cloud compute capabilities of Google Earth Engine (GEE) and Google Colab. This framework feeds a suite of remote sensing and data reanalysis inputs into a deep learning pipeline that deploys an artificial neural network to make pixel wise irrigation status determinations. We then inspect these outputs and compare them to currently available products to evaluate the strengths and differences among irrigation classification approaches.