Abstract: Autumn irrigation is one of the most critical water practices in the non-growing season, especially in seasonal freeze-thaw districts. Salt leaching and soil moisture preservation can be expected to serve for spring planting. This operation can historically account for approximately one-third of annual Yellow River water diversion in the Hetao Irrigation District of Inner Mongolia. However, the conventional extensive flood irrigation cannot fully meet the water regulations and volume reductions in sustainable agriculture. An urgent need is often required for a precise, efficient, and scalable monitoring framework for the spatiotemporal dynamics of autumn irrigation. In this study, an integrated framework was proposed and then validated for the high spatiotemporal resolution monitoring of autumn irrigation. Three sequential components comprised: preliminary extraction of water index, spatiotemporal data fusion, and fine classification with deep learning. The field test was conducted to extract the multi-source remote sensing data from the 2023 and 2024 irrigation periods in the Hetao Irrigation District. Three procedures were used: 1) Two complementary water indices were employed to enhance the spectral contrast of irrigation water and moist soil against bare land and crop residues. The multi-band water index (MBWI), which was incorporated green, red, near-infrared, and two shortwave infrared reflectance bands, was applied to moderate resolution imaging spectroradiometer (MODIS) and Landsat imagery for preliminary water detection. The normalized difference water index (NDWI) was applied to GaoFen-1 satellite data for fine boundary refinement and independent validation. A cropland mask from the China national land use and cover dataset often excludes natural water bodies and non-agricultural surfaces. 2) The object Landsat spatial and temporal adaptive reflectance fusion model (OL-STARFM) was implemented to reconcile the high temporal frequency of MODIS (500 m) with the spatial features of Landsat (30 m). The Landsat reference image was partitioned into spectrally homogeneous objects for the reflectance changes within each object. Synthetic cloud-free surface reflectance time series were generated at a 5-day interval and 30-meter resolution. The fusion was optimized for the regular field structure. Mixed-pixel effects were effectively avoided to obtain a consistent dataset for dynamic analysis. 3) An optimal multilayer perceptron (MLP) deep learning model was developed to classify irrigated versus non-irrigated fields. The architecture included an input layer, two hidden layers (128 and 27 neurons), and a binary output layer. Input features comprised multi-temporal spectral reflectance and derived MBWI time series. A five-fold data augmentation (random rotations, translations, and mirroring) enhanced the generalization. Training and validation were obtained from extensive field campaigns (60 sample points in 2023 and 1465 in 2024) and high-resolution unmanned aerial vehicle surveys with multispectral and thermal sensors. Results showed that the framework's effectiveness was achieved in balancing the strong cross-year stability and class performance from 2023 to 2024. Overall accuracy of classification reached 85.0% and 90.5%, respectively. Precision, recall, and F1-score metrics derived from confusion matrix analysis consistently exceeded 90% for both irrigated and non-irrigated classes. The cumulative autumn irrigation area was 288000 h in 2023 and 285600 h in 2024, respectively, which were concentrated in the Yichang, Jiefangzha, and Yongji sub-districts. McNemar's test and Kappa analysis indicated that statistically significant improvements were obtained in boundary delineation and temporal consistency over a single-source threshold. This framework can provide a reliable and practical solution for water-saving, irrigation scheduling optimization, and non-essential water reduction in the seasonal freeze-thaw regions, such as the Hetao Irrigation District.