A high dimensional features-based cascaded forward neural network coupled with MVMD and Boruta-GBDT for multi-step ahead forecasting of surface soil moisture

The objective of this study is to develop a novel multi-level pre-processing framework and apply it for multi-step (one and seven days ahead) daily forecasting of Surface soil moisture (SSM) based on the NASA's Soil Moisture Active Passive (SMAP)-satellite datasets in arid and semi-arid regions of Iran. The framework consists of the Boruta gradient boosting decision tree (Boruta-GBDT) feature selection integrated with the multivariate variational mode decomposition (MVMD) and advanced machine learning (ML) models including bidirectional gated recurrent unit (Bi-GRU), cascaded forward neural network (CFNN), adaptive boosting (AdaBoost), genetic programming (GP), and classical multilayer perceptron neural network (MLP). For this purpose, effective geophysical soil moisture predictors for two arid stations of Khosrowshah and Neyshabur were first filtered among 21 daily input signals from 2015 to 2020 by using the Boruta-GBDT feature selection. The selected signals were then decomposed using the MVMD scheme. In the last pre-processing stage, the most relevant sub-sequences from a large pool in previous process were filtered using the Boruta-GBDT scheme aiming to reduce the computation and enhance the accuracy, before feeding the ML approaches. The comparison of the results from the five hybrid and standalone counterpart models in term of standardized RMSE improvement (SRMSEI) revealed that M-VMD-B-o-C-PNN for SSM(T+1)| 27.13% and SSM (T+7)| 43.55% at Khosrowshah station and SSM(T+1)| 21.16% and SSM (T+7)| 30.10% at Neyshabur station outperformed the other hybrid frameworks, followed by M-VMD-B-o-B;-Ru, M-VMD-B-G-A(d.hooge), M-VMD-B-o-G(P), and M-VMD-B-o-M-LP. The accurately forecasted SSM data help improve irrigation scheduling, which is of significant importance in water use efficiency and food security.