Natural spring-based water potential in Bangladesh using ensemble machine learning algorithms

ABSTRACT The study introduces an advanced approach by incorporating ensemble machine learning algorithms to improve the overall accuracy of natural spring-based water potential in the Chittagong Hill Tracts (CHT) of Bangladesh. Among the 18 evaluated conditioning parameters, feature importance analysis revealed drainage density as the predominant one related to spring occurrence. Seven advanced ensemble models were applied to predict the spatial spring potentiality. Artificial neural network performed the best among them where AUC value of 0.93 and an accuracy of 0.77, reflecting its strength in modeling complex nonlinear relationships. According to the model outputs, high-potential zones occupy from 14.97 to 40.73%, while low-potential zones account for 1.17–7.29% of the CHT. District-wise outputs show marked variation as Khagrachari has the highest area of high potentiality at 41.52%, whereas Bandarban has the lowest, with 9.90%. According to future climate projections in different climate scenarios for 2030 to 2050, the results give an indication that parts of the southern areas of CHT could have greater potential due to high precipitation, while northern regions may experience low activity due to the decline in precipitation. These results offer crucial guidelines for sustainable water extraction with climate-resilient water management in the CHT.