Application of novel ensemble models and k-fold CV approaches for Land subsidence susceptibility modelling

Abstract

Land subsidence (LS) is significant problem that can lead to casualties, destruction of infrastructure, and socio-economic and environmental problems. In this study, we examine the Damghan Plain of Iran where LS poses a major obstacle to growth and management of the region. Dagging and random subspace (RSS) as meta- or ensemble-classifiers of a radial basis function neural network (RBFnn) were combined into two novel-ensemble intelligence approaches (Dagging-RBFnn and RSS-RBFnn) to predict and map the susceptibility of land units to subsidence. The goodness-of-fit (of training data) and prediction accuracy (of testing data) for the ensemble models were contrasted with the RBFnn, which is used as the benchmark for improvement. Details of 120 LS locations were examined and the data for twelve LS conditioning factors (LSCFs) were compiled. The LS points were randomly divided into four groups or folds, each comprised of 25 percent of the cases. The novel ensemble models were constructed using 75 percent (3 folds) and tested with the remaining 25 percent (onefold) in a four-fold cross-validation (CV) mechanism. Information-gain ratio and multicollinearity tests were used to select the LSCFs that would be used to estimate LS probabilities. The importance of each factor was calculated using a random forest (RF) model. The most important LSCFs were groundwater drawdown, land uses and land covers, elevation, and lithology. Twelve land subsidence susceptibility maps were generated using the k-fold CV approaches as each of the three models (RBFnn, Dagging-RBFnn and RSS-RBFnn) was applied to each of the four folds. The LS susceptibility models reveal a strong probability for LS on 15% to 24% of the plain. All of the maps generated achieved adequate levels of prediction accuracies and goodness-of-fits. The Dagging-RBFnn ensemble yielded the most robust maps, however. The ensemble of Dagging-RBFnn enhances the accuracy of modeling but the opposite condition was found for the RSS-RBFnn ensemble. It is evident that ensembles with meta classifiers might not always increase the accuracy of the base classifier. Overall, the southern part of the plain shows the highest LS risk. The results of this study suggests that groundwater withdrawal levels should be tracked and possibly restricted in regions with higher (extreme or moderate) probabilities of LS. This demonstrates that new approaches can support land use planning and decision making to minimize LS and improve sustainability.