Conventional and advanced AI-based models in soil moisture prediction

Abstract

This study evaluates the predictive accuracy of various computational models for Soil Moisture (SM) content, including (i) hard computing models such as mathematical (MLR) and stochastic techniques (AR, ARMA, ARIMA), and (ii) soft computing AI-based models such as shallow learning (NAR, NARX, MLPNN), advanced deep learning (LSTM, DRNN, CNN), and ensemble learning (Bagging, Boosting, and Adaboost). Using data from a sandy clay loam soil area, the models are developed and then compared for accuracy performance, tendency, and computational expense. Results indicate that hard computing models, particularly the stochastic AR model, do not act properly in predicting daily SM values, so that they could not improve the general predictive accuracy in comparison to the Naïve model based on several evaluation metrics. Shallow machine learning models like NAR and MLPNN perform better than the hard computing models, especially when they get the advantage of exogenous input vector (here, precipitation data). Deep learning models (Pearson Correlation Coefficient: PCC >0.9 and RMSE <1.39), especially the LSTM, exhibited higher accuracy than shallow learning models, however, they were the least favorite category in terms of computational cost. On the other hand, ensemble models show the best performance (PCC >0.92, RMSE >1.28) by combining multiple learners' strengths. In summary, the use of ensemble modeling improved the modeling accuracy of RMSE and PCC up to 6 % and 23 % in comparison to stochastic models, respectively.