Irrigation is the most water-consuming activity globally and continues to expand due to population growth and climate variability. Accurate quantification of irrigation water use (IWU) is essential for optimizing water resource allocation, especially in arid and semi-arid regions. However, existing remote sensing-based IWU estimation methods often underperform in arid regions due to data scarcity and insufficient representation of deep soil moisture dynamics. A novel method for estimating IWU was proposed by utilizing the similarities between heavy rainfall and irrigation events in arid regions. To better describe the nonlinear and complex relationships between irrigation and groundwater level, we firstly developed a machine learning model to overcome the difficulty of reproducing the spatiotemporal variability of specific yield in shallow aquifers. We then transferred the specific yield estimation model established during heavy rainfall to irrigation scenarios and derived the spatiotemporal distribution of IWU through the water balance equation. The proposed method was applied to the Hetao Irrigation District (HID), the largest gravity-fed irrigation district with single canal head in Asia. Our model accurately captured groundwater dynamics and yielded estimates of canal water use efficiency consistent with observations and previous studies. The IWU in the HID showed a declining trend, associated with the increase in sunflower cultivation area and the decrease in water diversion metrics. The spatiotemporal distribution of IWU exhibited a high degree of consistency with changes in cropping patterns. Our method reveals variations and intensity of human irrigation activity at high spatiotemporal resolution and provides important insights for water resource management in arid regions.