A hybrid clustering-fusion methodology for land subsidence estimation

Abstract

A hybrid clustering-fusion methodology is developed in this study that employs genetic algorithm (GA) optimization method, k-means method, and several soft computing (SC) models to better estimate land subsidence. Estimation of land subsidence is important in planning and management of groundwater resources to prevent associated catastrophic damages. Methods such as the Persistent Scatterer Interferometric Synthetic Aperture Radar (PS-InSAR) can be used to estimate the subsidence rate, but PS-InSAR does not offer the required efficiency and accuracy in noisy pixels (obtained from remote sensing). Alternatively, a fusion-based methodology can be used to estimate subsidence rate, which offers a superior accuracy as opposed to the traditionally used methods. In the proposed methodology, five SC methods are employed with hydrogeological forcing of frequency and thickness of fine-grained sediments, groundwater depth, water level decline, transmissivity and storage coefficient, and output of land subsidence rate. Results of individual SC models are then fused to render more accurate land subsidence rate in noisy pixels, for which PS-InSAR cannot be effective. We first extract 14,392 different input–output patterns from PS-InSAR technique for our study area in Tehran province, Iran. Then, k-means method is used to divide the study area into homogenous zones with similar features. The five SC models include adaptive neuro fuzzy inference system, support vector regression, multilayer perceptron neural network and two optimized models, namely radial basis function and generalized regression neural network. To fuse individual SC models, three methods including GA, K-nearest neighbors and ordered weighted average (OWA) based on ORNESS method and ORLIKE method, are developed and evaluated. Results show that the fusion-based method is significantly superior to each of the employed individual methods in predicting land subsidence rate.