Multi-timescale evapotranspiration fusion: A novel autoencoder with automated machine learning-based approach for enhanced estimation accuracy

Accurate quantification of evapotranspiration (ET) is critical for improving climate models, enhancing drought early warning systems, and optimising water resource management, as ET represents the largest water flux between land and atmosphere. However, precise ET estimation is often hindered by the complexity of data sources and temporal misalignments. To overcome these challenges, we proposed AGFusionET, a multi-timescale fusion model that integrates heterogeneous ET data from various sources, including remote sensing and climate models, to enhance ET estimation accuracy. AGFusionET uses AutoML and autoencoders to fuse data from 20 distinct ET products at fine temporal and spatial resolutions. Based on 585 eddy covariance datasets, we generated a long-term, high-resolution global ET dataset (0.05°) spanning 1982–2023, ensuring strong spatiotemporal continuity. The validation results show that AGFusionET outperforms all other benchmark ET products, achieving a Kling-Gupta Efficiency (KGE) of 0.88 and a Root Mean Square Error (RMSE) of 12.12 mm/month. AGFusionET excelled at both the monthly and annual scales at independent validation sites. By incorporating Normalised difference vegetation index (NDVI), Vapour pressure deficit (VPD), and diverse ET products that encode vegetation water status and irrigation-induced surface responses, AGFusionET partially reflects agricultural influences on ET even though explicit crop- and irrigation-related inputs are not included. Our study introduces a generalisable framework for multisource ET data fusion that provides more accurate and reliable ET estimates across diverse ecosystems, particularly in arid and high-latitude regions.