Examining the Potential Environmental Controls of Underground CO2 Concentration in Arid Regions by an SVD-PCA-ANN Preview Model

Abstract

This study attempts to examine environmental controls of the underground CO2 concentration, taking the CO2 concentration 4 m beneath the soil as an example. An SVD-PCA-ANN (singular value decomposition-principal component analysis-artificial neural network) preview model is proposed with the data of underground CO2 concentration and 12 environmental variables (the soil and meteorological data). The R2, RMSE, and RPD values of the proposed model are, respectively, 0.8874, 0.3351, and 2.7929, performing better than the popular preview models like SAE (stacked autoencoders), SVM (support vector machine), and LSTM (long short-term memory). It is proved that the underground CO2 concentration can be approximated by a nonlinear function of the considered variables. Soil temperature, salinity, and wind speed are the leading environmental controls, which explain 32.04%, 13.68%, and 11.21% in the variability of the underground CO2 concentration, respectively. Possible mechanisms associated with the environmental controls are also preliminarily discussed.

Machine learning brings new insights for reducing salinization disaster

“…The basic learning mechanisms of a neural network can be described as follows. The neuron input x was calculated from the model established in (Zhuang et al, 2021).…”

“…Differing from (Zhuang et al, 2021), we further introduce a braininspired mechanism in the learning processes. That is, we further improve SVD-PCA-ANN by long-short-memory neural network (LSTM) (Salman et al, 2018;Ali et al, 2020;Ahmad and Zhang, 2022;Gao, 2022;Rusnac and Grigore, 2022) and proposed a novel deep neural network PCA-ANN-LSTM to detect the potential environmental controls of soil salinity in the desert.…”

“…By testing the trained network, we can assess the errors between the outputs and correct results (Kato et al, 2016). Some latest researches constructed 3-stages learning systems (the training stage, validation stage and testing stage) to improve the generalization capability (Gatos et al, 2019;Skrede et al, 2020;Zhuang et al, 2021).…”