"Proposal for a Soil Erosion Prediction System in Peruvian Territory Using Satellite Imagery"

Abstract

"Soil erosion threatens agricultural and ecosystem stability in Peru, creating the need for efficient and scalable assessment tools beyond traditional monitoring. This research details the development of a soil erosion prediction system for the Piura region by integrating remote sensing, the RUSLE model, and deep learning. Our methodology involved two steps: first, building a multi-temporal geospatial database (2015-2024) on Google Earth Engine with 16 variables from Landsat 8, Sentinel-1, and MODIS. Second, developing two convolutional neural network architectures to predict the RUSLE A-factor. AU-Net model was used for direct spatial prediction, while a ConvL-STM U-Net model was designed for spatio-temporal forecasting. Evaluation in a test data set confirmed the high performance of both models. The U-Net accurately predicted the spatial distribution of erosion (R2=0.9895, MSE=0.0006, MAE=0.0021), and the ConvLSTM U-Net successfully forecasted its evolution (R2=0.7820, MSE=0.0036, MAE=0.0303). We conclude that this integrated system offers a powerful and scalable solution for erosion assessment. The ability to generate precise susceptibility maps and anticipate future behavior provides a valuable tool for land use planning and conservation strategies in vulnerable regions. © 2025 IEEE."