Wetland Habitat Health Condition Modeling Using Ensemble Machine Learning Algorithms

The construction of the Komardanga Dam on the Dhepa River at Birganj, Bangladesh, has precipitated significant ecological alterations in the downstream floodplain wetlands of the Punarbhaba River (Malda and Chapai Nawabganj of India-Bangladesh), marking a critical research problem. The substantial reduction in wetland area post-dam construction underscores the urgency to assess and model the health conditions of these habitats systematically. The major objective of this study was to leverage ensemble machine learningEnsemble machine learning models to evaluate wetland habitat health, aiming to generate a robust classification system that delineates varying health conditions across the wetland ecosystem. Methodologically, this research incorporated a suite of hydrological, water quality, and land composition indicators to inform the modeling process. Parameters such as water presence frequencyWater presence frequency (WPF), range of variability (RVA), and change rate of water presence were meticulously prepared, alongside indicators like agricultural presence frequency, Chlorophyll-a (Chl-a) concentration, and Temperature Condition IndexTemperature condition index (TCI) estimations derived from land surface temperatureLand surface temperature (LST) data. These indicators were integrated into ensemble machine learningEnsemble machine learning algorithms, including artificial neural networkArtificial neural network (ANN), support vector machineSupport vector machine (SVM), random forestRandom forest (RF), rotation forestRotation forest, and baggingBagging, to create a detailed model of wetland habitat health. The models’ accuracy was rigorously validated using empirical and binormal ROC curves. Quantitative results revealed a stark decline in wetland areas, from 255.84 km2 in 1993 to 72.2 km2 by 2019. Indicators such as NDWINormalized difference water index and WPFWater presence frequency indicated healthier conditions in central wetland areas, in contrast to stressed peripheral zones. According to the random forest model, only 13.75% of the wetland area is classified as ‘very good’, while a substantial 54.8% is deemed ‘poor’ or ‘very poor’. Comparatively, the ANN model identifies 13.31% as ‘very good’ and a combined 45.7% as ‘poor’ or ‘very poor’. The ensemble machine learningEnsemble machine learning models identified the most degraded regions, with the rotation forestRotation forest algorithm displaying superior performance with a binormal ROC AUC of 0.936. The SVM and baggingBagging models also showcased high accuracy, with binormal AUCs of 0.93 and 0.92, respectively, while the ANNArtificial neural network and RF models demonstrated competent performance with binormal AUCs of 0.912 and 0.873. This study’s integrated approach, combining detailed parameter analysis with sophisticated machine learning algorithms, offers a comprehensive tool for environmental monitoring and underscores the need for strategic wetland management to mitigate the adverse effects of hydraulic structures on these critical ecosystems.