A novel hybrid machine learning framework for spatio-temporal analysis of reference evapotranspiration in India

The study focuses on the diverse climatic regions of India, spanning arid, semi-arid, sub-humid, and humid zones. This research employs a novel hybrid machine learning (ML) framework for precise spatio-temporal reference evapotranspiration (ET o ) modelling from 1970 to 2024, addressing the variability in temperature, humidity, and precipitation. Three advanced ML models—Quantile-Adjusted xLSTM Network (QAxLNet), Quantile-Score Diffusion Model (QSDM), and Attentive Deep Quantile-Aware Autoencoder Network (ADAQNet)—are proposed and applied, focusing on relative humidity and temperature as critical predictors. Model validation, conducted with EEFlux-derived ET o data and Indian Meteorological Department (IMD) benchmarks, revealed strong alignment across diverse climatic zones. The MPI-ESM1–2-HR model under SSP3–7.0 scenarios outperformed other CMIP6 models, with a correlation coefficient of 0.975 and spatial error of 3.55 mm. The ADAQNet demonstrated superior performance, with lowest errors RMSE (Train: 0.2247, Test: 0.2499), and R 2 (Train: 0.96; Test: 0.9571) among the models. ET o declined at an average rate of 1.9 mm/year, indicating the role of climate change. ET o variability closely mirrored the spatial distribution of the National Building Code (NBC) of India. Seasonal variations were significant, with arid regions (Rajasthan, Gujarat) experiencing the highest increase (2.5–5.1 mm/year). Humid regions showed high sensitivity to RH forecasts, with up to 20 % ET o deviation. The study emphasizes the spatial, temporal, and seasonal variations of ET o across the region, highlighting its dependence on climatic factors. • CMIP6 model MPI-ESM1-HR with SSP3–7.0 represents the Indian climate scenario best. • Key predictors for RH downscaling selected from ERA5 factors using SHAP values. • Proposed hybrid ML model ADQANet was most accurate statistical downscaling model. • Downscaled RH aligns with NBC climate zones, with minimal errors of 2.1 % for RH max . • Calculated ET o matches with the EEFlux data with an error range of −1.2 % to 2.5 %.