Hybrid Tree-Based Wetland Vulnerability Modelling

Abstract

Wetlands of the moribund region of the Ganga–Brahmaputra deltaic part experience extreme loss and degradation, which is the leading cause for our present study. In this study, the vulnerable situation, as a part of degradation, is explored using tree-based ML algorithms in python environment using eight conditioning parameters, namely: water presence frequency (WPF), change in WPF, hydro duration, water depth, agriculture presence frequency, proximity to the river, distance from the road network, and built-up proximity. Four tree-based machine learning algorithms, namely, bagging classification model, reduced error pruning tree (REP Tree), gradient boosting classification model (GBM), and AdaBoosting classification model (ADB), has been used to evaluate the vulnerability of wetlands for both phase II (1998–2007) and phase III (2008–2017). It is found that 23.92–25.01% and 44.67–46.99% area to total wetland area emerged as high to very high vulnerable zone in phase II, whereas 24.08–26.16% and 45.41–49.13% of wetland area identified as high to very high vulnerable zone in phase III. More than 45% of the total wetland area disappeared during phase II to phase III. The models have been validated using the following matrices like sensitivity, Precision F1-score, and MCC for justifying the best-suited model. With an average score of more than 91 for all the matrices, the gradient boosting classification model (GBM), and AdaBoosting classification model (ADB) exhibit more prediction capability and model accuracy than the bagging classification model, and Reduced Error Pruning (REP) Tree model. With the successful prediction, the study recommends tree-based ML algorithms for such or similar works. The study also warns about growing wetland habitat vulnerability and its negative consequences on socio-ecological benefits.