Abstract: Autumn irrigation is a crucial non-growing season water management practice in seasonal freeze-thaw zones for salt leaching and moisture preservation, accounting for approximately one-third of the annual irrigation water use. In the context of increasingly stringent water resource constraints, the traditional flood irrigation approach has become unsustainable, urgently necessitating the development of accurate and efficient techniques for monitoring and managing autumn irrigation. This study, focusing on the Hetao Irrigation District in Inner Mongolia, presents a novel, stepwise technical framework integrating preliminary water index extraction, spatiotemporal data fusion, and deep learning-based fine classification to achieve high spatiotemporal resolution remote sensing monitoring of the autumn irrigation process. The framework first employs the Multi-band water index (MBWI) and the Normalized difference water index (NDWI) to preliminarily identify irrigation water signals from multi-source remote sensing data. It then utilizes the Object-based landsat spatial and temporal adaptive reflectance fusion model (OL-STARFM) to fuse MODIS and Landsat imagery, generating a high-quality time-series dataset at 5-day intervals with a 30-meter spatial resolution. Finally, an optimized Multilayer perceptron (MLP) model is applied to perform automated, fine-scale classification of the irrigated areas. Results from the 2023 and 2024 monitoring periods demonstrate overall accuracies of 85.0% and 90.5%, respectively, with recognition accuracy for both irrigated and non-irrigated fields consistently exceeding 90%. This validates the method's effectiveness and robustness under complex surface conditions. The study provides a reliable technical approach and data support for water-saving management, irrigation scheduling, and the identification of non-essential water use in the Hetao Irrigation District and similar seasonal freeze-thaw irrigation areas.