Proceedings Article | 13 September 2024
KEYWORDS: Soil science, Spatial resolution, Data modeling, Data archive systems, Artificial intelligence, Temperature metrology, Sand, Environmental monitoring, Deep learning, Data fusion
Improved management of grazing resources has proven to be effective in mitigating soil erosion and enhancing carbon sequestration. Efficient monitoring of soil descriptors plays a crucial role in achieving this goal, as it provides valuable information for evaluating soil loss estimation by water erosion based on the Revised Universal Soil Loss Equation (RUSLE) model. The accuracy of RUSLE model depends on the quality of the input soil data, namely, soil texture and organic carbon. However, the existing soil spatial products are created using conventional machine learning methods, which combine spaceborne spectral input data with environmental covariates, resulting in moderate performance and coarse resolution. Therefore, novel approaches are needed to tackle the challenge posed by the synergistic framework of data analytics, which require effective fusion of multispectral data with environmental and topographical covariates. In this study, we explore the potential of employing a deep learning architecture to obtain a new data representation from spaceborne Sentinel-2 information for the regression task. Concurrently, we feed an eXtrem Gradient Boosting (XGBoost) regressor, with the (128) features extracted by a convolution neural network (CNN). Additionally, 85 spatial layers, representing landscape features, and bioclimatic variables, have also been used as input features in the XGBoost regressor. The CNN-XGBoost model was trained using a subset of 83 Greek soil samples corresponding to grassland from the LUCAS 2015 dataset. The generation of enhanced soil input layers, including clay and organic carbon, resulted in a reduction of RMSE. These spatial products were integrated into the RUSLE to improve the soil erodibility factor, leading to the creation of a soil erosion layer with higher spatial resolution (10m). Mapping conducted at a study site with significant areas of grasslands in Elassona, Greece, highlight the importance of our approach compared to existing soil products.
