Machine learning approaches for the prediction of soil aggregate stability

Currently, many Pedotransfer Functions (PTFs) are being developed to predict certain soil properties worldwide, especially for difficult and time-consuming parameters to measure. However, very few studies have been done to assess the feasibility of using PTFs (regression or machine learning methods) for predicting soil aggregate stability. Also, the Random Forest (RF) method has never been used before to predict this parameter, and no study was found concerning the use of PTFs methods to estimate soil parameters in Morocco. Therefore, the current study was conducted in the three watersheds of Settat- Ben Ahmed Plateau, located in the center of Morocco and covering approximately 1000 km2. The purpose of this study is to compare the capabilities of the machine learning technique (Random Forest) and Multiple Linear Regression (MLR) to predict the Mean Weight Diameter (MWD) as an index of soil aggregate stability using soil properties from two sources data sets and remote sensing data. The performance of the models was evaluated using a 10-fold cross-validation procedure. The results achieved were acceptable in predicting soil aggregate stability and similar for both models. Thus, the addition of remote sensing indices to soil properties does not improve models. Results also show that organic matter is the most relevant variable for predicting soil aggregate stability for both models. The developed models can be used to predict the soil aggregate stability in this region and avoid waste of time and money deployed for analyses. However, we recommend using the largest and most uniform possible data set to achieve more accurate results. Pedotransfer functions, soil aggregate stability, mean weight diameter, multiple linear regression, random forest, remote sensing data.