Hydro-environmental dynamics of Kaptai Lake using satellite derived biophysical metrics and an ensemble Machine...

Bangladesh is a land of numerous fluvial waters bodies traversing across the whole landscape which has the complex waterbodies form the hydrological and ecological regime. Prediction of watershed change intrinsically imposes challenges due to complex hydro-climatic reasons and water quality processes. The Kaptai lake manifests fluctuations for its seasonal precipitation and changes in vegetation condition. The remote sensing-based watershed behavior analysis using Machine learning (ML) architecture particularly Stacked Regression (SR) for Kaptai boundary is seen to have a potential to address the variability that is absent in the current literature. This study has been conducted using the Kaptai watershed, which is one of the largest artificial lake of South-East Asia, with the objective for developing a data driven methodology for a sustainable forecasting of watershed behavior. Thus, water Quality Index, meteorological variables, integrative synthesis of remote sensing indices along with combination of Machine Learning Architecture have been applied. For capturing the spatial and temporal dynamics, spatial maps spanning the 1990-2025, a period of 35 years have been generated using ArcMap (ArcGIS) for salient biophysical and water quality parameters. The study reveals that the steep slope of the Chittagong Hill Tracts (CHT) is the root cause of high TDS and Turbidity (NDTI). The heavy rains and land-use changes, namely deforestation and shifting cultivation further assist in transport of large sediment load into the reservoir. The indication of a positive non-linear water quality dynamics, periodic eutrophication and turbidity peaks (NDCI up to 0.99) and spikes, and significant vegetation loss and significant increase in LST are seen to play the role. A LightGBM model for Time Series prediction and a Stacked Regression model comprising of XGBoost, Random Forest and Multi-Layer Perceptron has been integrated through RidgeCV meta-Learner. The derived analytical outcomes give an R² = 0.977 from Stacked Regression modelling. Moreover, by utilizing Light GBM and ANN an R² value of 0.9402 and 0.99 have been obtained, respectively. This framework shows substantive functionality for forecasting in a scalable approach by involving hydrological, and biophysical processes within the watershed under complex variability. The findings here are a pathway towards an empirically grounded water management and climate resilient behaviors of Kaptai lake, that in turn help in holistic decision making.