Study region: The upper reaches of the Yangtze River (URYR) of China Study focus: This study investigates the spatio-temporal variations of hydrological cycle components in the upper reaches of the Yangtze River (URYR) from 1980 to 2030 using multi-source remote sensing data and machine learning. A long short-term memory network with an attention mechanism (LSTM-AT) is initially developed to predict exploitable water resources (EWR) in the URYR. New hydrological insights for the region: Significant spatial and temporal variations are identified over the past four decades (1980-2015) and projected through 2030. Precipitation ranges from about 230 mm in the arid northwest to over 1600 mm in the humid southeast, with evapotranspiration varying from less than 100 mm at high altitudes to over 800 mm in low-lying areas. Runoff and terrestrial water storage exhibit similar spatial gradients, and runoff shows strong dependence on precipitation, particularly in the Min-Tuo River basin. LSTM-AT predictions suggest that the regional water balance remains dynamically stable, while EWR in the upper Jinsha River shows a fluctuating upward trend of 2.13 +/- 0.87 mm/yr, underscoring the need for adaptive water resource management under ongoing environmental change. By leveraging an attention-based deep learning framework and interpretable feature-weight analysis, this work also provides new insight into the dominant hydro-climatic and human-regulation drivers shaping annual water availability in this highly regulated basin.