3 - New formulation for predicting soil moisture content using only soil temperature as predictor: multivariate adaptive regression splines versus random forest, multilayer perceptron neural network, M5Tree, and multiple linear regression

Abstract

Soil moisture (SM) content is an important component of the soil water budget and plays a critical role in agricultural, hydrological, and water resources management, especially in the determination of crop water requirement. SM content can be determined using direct or indirect methods. In the present investigation, we propose a new method for SM estimation using artificial intelligence models, by linking SM to soil temperature. This chapter shows that using only soil temperature as predictor, SM can be calculated with very high accuracy and precision. Also, it is demonstrated that depending on the periodicity represented by the component of the Gregorian calendar, for example the year numbers, month numbers, day numbers, and hour numbers, as predictors, the proposed models can lead to high accuracy. In particular, we perform a comparison between four machine learning models, namely, multivariate adaptive regression splines (MARS), random forest (RF), M5Tree, and the multilayer perceptron neural network (MLPNN). Results obtained using the machine learning models were compared to those obtained using the multiple linear regression (MLR) model. All the proposed models were applied and compared using data collected at two stations operated by the United States Geological Survey (USGS). The accuracy of the models was evaluated using coefficient of correlation (R), Willmott’s index of agreement (d), root mean squared error, and mean absolute error. First, the models were developed using only the soil temperature as input variable. Obtained results show moderate-to-low accuracy. Second, the periodicity was included as input variable combined with the soil temperature, which leads to significant improvement in models performances, and the models become more robust and accurately estimate the SM. In contrast, the standard MLR method was not able to provide high accuracy, and the performances of models were slightly improved. For the USGS 01315226 station, the MARS model provides the best accuracy with significant R and d values of 0.967 and 0.983, on equal terms with the RF model (R=0.965, d=0.981), higher than the MLPNN model (R=0.941, d=0.969), and the M5Tree model (R=0.962, d=0.981), and much better than those obtained using the MLR model (R=0.580, d=0.488). While for the USGS 01315227station, the RF model provides the best accuracy with significant R and d values of 0.992 and 0.996, higher than the MARS model (R=0.988, d=0.993), higher than the M5Tree model (R=0.941, d=0.969), and the MLPNN model (R=0.978, d=0.989), and much better than those obtained using the MLR model (R=0.294, d=0.389). Overall, the proposed approach will be helpful for SM estimation.