Developing Surrogate Soil Erosion Models for Conservation Program Delivery

The USDA Farm Service Agency currently ranks its offers to landowners within the Conservation Reserve Program CRP by estimating environmental benefits that include water/wind erodibility indices using climate and soil factors associated with legacy models, such as RUSLE and WEQ. However, unit-less index values reflect soil loss reduction much less clearly than actual erosion rates provided by model simulations using contemporary WEPP and WEPS erosion models, developed at the USDA Agricultural Research Service. They provide the latest in erosion science. However, their lengthy simulation run-times may inhibit direct use of those process models in conservation program application ranking workflows. During heavy sign-up periods the calculation of environmental benefits, including erosion control, should take a few seconds per application. Recent machine learning advances facilitate the development of surrogate models (SM)s, potentially providing sufficiently accurate and much faster results. They are models of models. Their adoption requires understanding the tradeoffs between response times, complexity, and acceptable output accuracy. CRP is a suitable application for such models. Here, the Cloud Services Integration Platform (CSIP) was extended and utilized to automatically obtain data from full erosion model simulations and derive the erosion SMs at the modeling framework level. NeuroEvolution of Augmenting Topology (NEAT) techniques in an ensemble application, combined with Artificial Neural Networks (ANN) uncertainty quantification and other approaches are the main methodologies used. We compare soil erosion control benefits calculated by WEPP and WEPS SMs to full model runs. The SMs are ANN ensembles, generated from large training and validation datasets running full model simulations for fallow conditions on a ~900-meter grid across the program landscape, and deployed through a web service. The full WEPP and WEPS models are also deployed through a web service. Both services types fetch soil and climate input data from common USDA on-line sources. The SM service meets the few-second threshold with most run-time spent fetching input data, and the SMs themselves taking milliseconds to finish. Results align well with full simulations (Nash-Sutcliffe model efficiencies > 0.95) for given study areas.