Estimation of soil properties using machine learning techniques to improve hydrological modeling in a semiarid environment: Campo de Cartagena (Spain)

Abstract Soils are a key element in the hydrological cycle through a number of soil properties that are complex to estimate and exhibit considerable spatial variability. Therefore, several techniques have been proposed for their estimation and mapping from point data along a given study area. In this work, four machine learning methods: Random Forest, Support Vector Machines, XGBoost and Multilayer Perceptrons, are used to predict and map the proportions of organic carbon, clay, silt and sand in the soils of the Campo de Cartagena (SE Spain). These models depend on a number of hyperparameters that need to be optimised to maximise accuracy, although this process can lead to overtraining, which affects the generalisability of the models. In this work it was found that neural networks gave the best results in validation, but on the test data the methods based on decision trees, random forest and xgboost were more accurate, although the differences were generally not significant. Accuracy values, as usual for soil variables, were not high. The RMSE values were 8.040 for SOC, 7.049 for clay, 10.227 for silt and 13.561 for loam. The layers obtained were then used to obtain annual curve number layers whose ability to reproduce runoff hydrographs was compared with the official CN layer. For high flow events, the CN layers obtained in this study gave better results (NSE=0.807, PBIAS=-4.7 and RMSE=0.4) than the official CN layers (NSE=-2.28, PBIAS=135.82 and RMSE=1.8).