Deep Learning Approach for Water Erosion Mapping Using Google Earth Engine

Soil loss by water erosion is a major risk affected a food security in the world and especially in the Mediterranean region. In Tunisia, this phenomenon is accentuated by soil, land use/cover characteristics. The assessment of soil loss using a physically-based, distributed, and continuous model is very important, but it is also highly requested by daily and high resolution data. These tasks are very difficult for human analysis. Artificial Intelligence and a subtype of machine learning called deep learning is used to perform with more speed, consistency and perhaps more accurate than humans can perform. This study is focused on M’richet El Anse watershed, characterized by a moderate Mediterranean climate and a high spatial and temporal heterogeneity of land use proprieties. We used the Google Earth Engine to execute a deep learning algorithm to extract, temperature, solar radiation, land use, land cover change, and rotation from MODIS image from 01/01/1994 to 31/08/2002. The outputs of ANSWERS-2000 model were evaluated by comparison between the predicted and observed values using statistical coefficients including coefficient of determination and Nash–Sutcliffe efficiency. Results show that the model outperformed the usually estimated input parameters in the assessment of the annual soil erosion (R-squared: 0.81 and NSE: 0.72). Finally, the use of deep learning for big spatial data makes the use of the hydrological and water erosion model, ANSWERS-2000, more and best adopted for the estimation, quantification, and spatial variation of water erosion at the watershed scale.