LSTM integrated with Boruta-random forest optimiser for soil moisture estimation under RCP4.5 and RCP8.5 global warming scenarios

• Published in: Stochastic environmental research and risk assessment Vol. 35; no. 9; pp. 1851 - 1881
• Main Authors: Ahmed, A. A. Masrur, Deo, Ravinesh C., Ghahramani, Afshin, Raj, Nawin, Feng, Qi, Yin, Zhenliang, Yang, Linshan
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.09.2021; Springer Nature B.V
• Subjects: Algorithms; Aquatic Pollution; Chemistry and Earth Sciences; Climate change; Climate models; Computational Intelligence; Computer Science; Deep learning; Earth and Environmental Science; Earth Sciences; Environment; Environmental health; Environmental management; Feature extraction; Feature selection; Global climate; Global climate models; Global warming; Hydrology; Hydrometeorology; Learning; Long short-term memory; Machine learning; Math. Appl. in Environmental Science; Original Paper; Physics; Probability Theory and Stochastic Processes; Soil dynamics; Soil moisture; Statistics for Engineering; Waste Water Technology; Water Management; Water Pollution Control; Global climate model (GCM); Hybrid model; Murray darling basin; Soil moisture; Boruta-random forest optimiser algorithm (BRF); Long short-term memory (LSTM)
• ISSN: 1436-3240; 1436-3259
• DOI: 10.1007/s00477-021-01969-3

Future soil moisture (SM) estimation is a practically useful task for eco-hydrologists, agriculturists, and stakeholders in environment health monitoring to generate comprehensive understanding of hydro-physical and soil dynamic system. This paper demonstrates the capability of a hybridised long short-term memory (LSTM) predictive framework to emulate SM under global warming scenarios. The proposed model is developed by integrating Boruta-random forest (BRF) feature selection and capturing significant antecedent memory of SM behaviour were applied to estimate the future SM using Coupled Model Intercomparison Phase-5 (CMIP5) repository. The BRF is adapted to extract pertinent features in hydro-meteorological variables intrinsically related to SM, and therefore, is used to construct a hybridised deep learning (i.e., BRF-LSTM) model. To establish the viability of deep learning model for SM estimation until 2100, five stations closely matched to the global climate model grid are selected in Australia's Murray Darling Basin. The performance skill of BRF-LSTM model is compared against standalone models (i.e., LSTM, SVR, and MARS). The results showed that the hybrid deep learning model (i.e., BRF-LSTM) with a feature selection capability could significantly outperform the standalone models for both warming simulations. The proposed hybrid model also demonstrated superiority in SM estimation with over 95% of all predictive errors lying below 0.02 mm, and low relative root means square error (≈ 1.06% for RCP4.5 and ≈ 1.888% for RCP8.5) to outperform all the benchmark models. This study demonstrates the capability of LSTM algorithm coupled with BRF feature selection to simulate future soil moisture under climate change, and so, can be successfully implemented in hydrology, agriculture, soil use management and environmental management.