ASABE Technical Library - Abstract

Estimating Soil Erosion for Large Watersheds Using Machine Learning and RUSLE2

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  Soil Erosion Research Under a Changing Climate, January 8-13, 2023, Aguadilla, Puerto Rico, USA  .(doi:10.13031/soil.23084)
Authors:   Dalmo A. N. Vieira, Brandon W. Sims, Robert R. Wells, Daniel C. Yoder, Ronald L. Bingner
Keywords:   Artificial neural networks, Erosion modeling, Machine learning, RUSLE2, Soil erosion.

Abstract

A new modeling framework has been developed to extend RUSLE2 erosion prediction technology for large areas with the use of machine learning and improved geoprocessing tools. RUSLE2 has been traditionally used to estimate soil loss over one-dimensional hillslopes that would be representative of the erosion for an area. The new approach calculates sheet-and-rill erosion in two horizontal dimensions, producing detailed maps of soil loss that can be used to identify critical areas and guide the design of soil conservation measures for entire watersheds. Geoprocessing algorithms developed for this application utilize high-resolution gridded digital elevation models to analyze overland flow paths and create a drainage network and corresponding hillslopes, ensuring the correct representation of topography and runoff distribution for RUSLE2 calculations. Topographic attributes, soil types, and land management parameters are extracted for each hillslope from GIS layers and associated databases. Soil erosion is then inferred from a machine learning procedure that uses a sequential, densely connected artificial neural network (ANN). A total of 12 input parameters describe climate, topography, soil properties, and vegetation and agricultural management operations for each hillslope. The ANN determines the corresponding long-term average soil loss of the area. To illustrate the methodology, a map of average annual soil loss for the entire state of Iowa, USA, was created covering all areas typically planted with corn and soybeans that were assumed to be managed as two-year rotations, with a conservation tillage practice, which cover about 96% of the state. Elevation data was derived from the state‘s Lidar survey, resampled to 10-meter resolution. Land use data was obtained from the USDA-NASS Crop Data Layer for 2018. A set of independent, simple-profile RUSLE2 calculations were used to train the ANN. The training set included five climate definitions for different regions of the state, 7 soil types from the SSURGO database of varying erodibility, and 10 typical RUSLE2 management descriptions for corn and soybean rotations with varying yields. Topography was represented by uniform hillslopes with lengths varying from 25 to 300 ft (7.6 to 91 m) and slopes between 1% and 18%. All input parameters were combined to create a training set of 459,900 simulation results. Validation tests when compared with average values calculated with the original RUSLE2 model showed that ANN computed erosion values were within 0.27 tons per acre (0.61 Mg/ha). Application of the geoprocessing algorithms for the entire state required 20 hours of processing time across 12 cores using Message Passing Interface parallelization. The calculation of soil loss through the trained ANN for about 186 million hillslopes in 1712 HUC-12 watersheds covering the entire state was computed in about 30 minutes on a single core. A similar calculation using RUSLER-Distributed on a 192-core server (websim.rusle2.org) would take about 43 days. The new methodology shows that the introduction of machine learning allows RUSLE2 to be easily extended to watershed scales while maintaining a high level of spatial detail, making it a useful tool for prioritization of areas for erosion control, evaluating the impact of best management practices, or in the design of soil conservation practices.

(Download PDF)    (Export to EndNotes)