Structural Reorganization of Water Use Efficiency Drivers in the Hanjiang River Basin Under the South-to-North Water Diversion: Spatiotemporal Deep Learning and Pixel-Level Attribution

Interbasin water transfer reshapes water use efficiency (WUE) by altering hydrology and land use, yet post-transfer WUE changes and driver reorganization remain poorly quantified. Focusing on the Hanjiang River Basin (South-to-North Water Diversion source), we employed a convolutional gated recurrent unit (ConvGRU) deep learning model with Integrated Gradients to quantify pixel-scale spatiotemporal contributions of nine drivers to WUE over 2000–2024 using meteorological and multi-source remote sensing data. Results show (1) Basin-wide WUE increased over 2000–2024, but post-diversion growth dropped sharply (0.0184 to 0.0045 yr−1), with river-channel declines driving this deceleration. (2) Temperature and vegetation land-cover type were the strongest suppressive drivers, while cropland and vapor pressure were the strongest promotive drivers; after diversion, kNDVI exhibited the largest attribution increase (+0.14), and the center of gravity of driving forces is shifting from climatic factors toward vegetation and land-use factors. (3) Across the basin, 42.26% of areas underwent a switch in dominant driver, with the most intense reorganization occurring in the upper reaches; the dominant area of cropland rose from 26.28% to 29.24%, kNDVI nearly doubled from 9.09% to 16.64%, and its expansion simultaneously absorbed dominance from temperature, radiation, cropland, and vegetation land-cover type—a “multi-source convergence” pattern of reorganization. Sensitivity analysis across multiple baseline schemes confirmed the robustness of the attribution conclusions. These findings inform ecological and agricultural water management in the Hanjiang Basin under the diversion.