Reconstructing missing time-varying land subsidence data using back propagation neural network with principal component analysis

Abstract

Abstract Land subsidence, a complex geophysical phenomenon, necessitates comprehensive time-varying data to understand regional subsidence patterns over time. This article focuses on the crucial task of reconstructing missing time-varying land subsidence data in the Choshui Delta, Taiwan. We propose a novel algorithm that leverages a multi-factorial perspective to accurately reconstruct the missing time-varying land subsidence data. By considering eight influential factors, our method seeks to capture the intricate interplay among these variables in the land subsidence process. Utilizing Principal Component Analysis (PCA), we ascertain the significance of these influencing factors and their principal components in relation to land subsidence. To reconstruct the absent time-dependent land subsidence data using PCA-derived principal components, we employ the backpropagation neural network. We illustrate the approach using data from three multi-layer compaction monitoring wells from 2008 to 2021 in a highly subsiding region within the study area. The proposed model is validated, and the resulting network is used to reconstruct the missing time-varying subsidence data. The accuracy of the reconstructed data is evaluated using metrics such as root mean square error and coefficient of determination. The results demonstrate the high accuracy of the proposed neural network model, which obviates the need for a sophisticated hydrogeological numerical model involving corresponding soil compaction parameters.