Reliable basin-scale projection of evapotranspiration at fine spatiotemporal scale using machine learning-based techniques

Accurate evapotranspiration (ET) projections are essential for water resource management and drought prediction under a changing climate. However, ET estimates from Global Climate Models (GCMs) exhibit large uncertainties stemming from systematic biases in meteorological variables, coarse spatial resolution, and inter-model structural differences. This study addresses these limitations by developing an integrated framework combining CMIP6 projections, a proxy for observations, and a machine learning approach to improve the reliability of projected ET at a fine spatiotemporal scale (temporal: daily scale; spatial: 0.1°×0.1°) over the Krishna River Basin. To select a proxy for observed multiple ET products (ERA5-Land, GLDAS-NOAH, and GLEAM) were first validated against the water balance-based basin scale ET available from Ma et al. (2024). Results indicate superior performance of ERA5-Land for the selected river basin and was subsequently used to assess the uncertainty of seven CMIP6 GCMs (over the historical period 2015–2023). Through multi-metric analysis, EC-Earth (SSP5-8.5) outperformed all other models, showing low weighted mean absolute error (WMAE,) highest correlation and strongest 1:1 alignment. Despite this, initial evaluation of raw EC-Earth inputs revealed systematic biases in variables associated with ET like solar radiation (ssrd), thermal radiation (strd), and sensible heat flux (sshf), leading to substantial underestimation of ET. To resolve this, a Quantile Mapping (QMAP) technique was employed to bias-correct the meteorological drivers, successfully restoring their statistical distributions. Then a data-driven Bayesian Network (BN) model was developed to simulate ET using precipitation, temperature, and the bias corrected variables. The BN model demonstrated robust performance (R > 0.87) during both development and testing phases. Near-future projections (2024–2030) indicate that relying on raw GCM data dampens seasonal cycles; in contrast, the bias-corrected BN projections highlight a higher mean ET and effectively capture seasonal extremes, particularly during dry months (January–June). These findings underscore the critical role of bias correction in hydro-climatic modelling and establish this framework as a reliable tool for future hydrological assessment in the Krishna River Basin. This integrated methodology demonstrates that coupling statistical bias correction with machine learning models can substantially reduce projection uncertainty. The framework is transferable to other basins and provides reliable ET projections for improved water availability assessments and climate adaptation planning.

Keyword: - Quantile Mapping, Bayesian Network, CMIP6 downscaling ,Bias correction, Hydro-climatic modelling.