Abstract: Accurate identification of irrigated farmland is crucial for optimizing water resource allocation, improving irrigation efficiency, monitoring agricultural drought, guiding hydraulic engineering operations, and promoting food security and sustainable water use. This study focuses on the entire region of Uxin Banner, Inner Mongolia, an arid–semiarid agro-pastoral ecotone. Differences between irrigated and non-irrigated farmland are compared and examined from four dimensions－moisture dynamics, energy dynamics, vegetation physiology, and vegetation phenology－leading to 10 key factors selected across four categories. We retrieve these factors from multi-source remote sensing data, and construct a comprehensive feature index library comprising 43 indicators. Then, we select the optimal feature subset through feature-importance evaluation, and use it to train a Random Forest (RF) model for irrigated farmland identification. Application to the 2019-2024 Uxin Banner data for analysis of the spatiotemporal variations shows that the model achieves an overall accuracy of 95.38% and a Kappa coefficient of 0.9128, demonstrating its strong capability of irrigated farmland extraction. In Uxin Banner, the irrigated area was 812.2733 km2 in 2024, mainly distributed in its southern and southwestern regions. Over the five years of 2019-2024, the irrigated area increased first and then decline with a turning point occurring in 2022. The new framework in this study—mechanism analysis, factor identification, remote sensing inversion, and model identification—would help identify irrigated farmland and promote interpretable machine learning applications.

Key words: irrigated farmland, multi-source remote sensing, mechanism-based analysis, random forest, accurate identification