Abstract: Maize is one of the most important food crops in the world. It is often required to timely and accurately acquire its spatial distribution for the high yield and national food security in sustainable agriculture. This study aims to precisely and efficiently extract the large-scale maize planting areas using deep learning and Sentinel-2 Image data. The research area was taken from the Santun River Irrigation District in Xinjiang, China. Firstly, the field surveys were provided by the Santun River Basin Management Office. The optimal period of extraction was identified as July to August, with the peak spectral characteristics of the maize canopy at the growth stage. Secondly, the primary data source was used as the Level-2A surface reflectance product of Sentinel-2 satellite imagery from 2019 to 2024. A series of preprocessing steps—including resampling with uniform spatial resolution, band fusion to integrate spectral information, mosaicking to cover the entire irrigation district, and cropping to the study area boundary—were performed to obtain 10-meter resolution multispectral Sentinel-2A imagery of the irrigation district. Thirdly, the field survey data were collected during the growing seasons. Ground truth labels were manually annotated for maize. A robust dataset was constructed for training deep learning networks. Data augmentation techniques were then applied to enhance dataset diversity and model generalization, such as symmetric transformation, rotation at various angles, and brightness adjustment. The dataset size was effectively expanded to reduce the overfitting risks. A CBAM-UNet model with the convolutional block attention module (CBAM) was then established on the Pytorch platform. Three widely used deep learning models—U-Net, SegNet, and DeepLabV3+—were selected as the benchmarks to evaluate the performance of the CBAM-UNet model. Finally, multiple evaluation metrics were employed, including mean intersection over union (mIoU), mean pixel accuracy (mPA), F1-score, overall accuracy, absolute error (AE), absolute percentage error (APE), and coefficient of determination (R²). The accuracy of each model was assessed to extract the maize planting areas against ground truth data and official statistical records. The results indicate that: 1) The CBAM-UNet model achieved the highest accuracy of extraction among all models, with an mIoU of 83.14%—representing significant improvements of 1.93, 9.57, and 4.96 percentage points over U-Net, SegNet, and DeepLabV3+, respectively. The mPA reached 87.84%, with the increases of 1.97, 4.46, and 3.82 percentage points, compared with the remaining three models. The F1-score was 89.62%, which was 1.58, 6.39, and 2.82 percentage points higher than the rest, respectively. The overall accuracy was 96.80%, exceeding by 1.24, 4.12, and 2.89 percentage points, respectively. 2) The absolute error (AE) between the maize area extracted by the CBAM-UNet model and official statistics from the Santun River Basin Management Office ranged from 406.01 to 796.20 hectares over the six-year period, indicating the small and concentrated errors with the minimal deviation. All absolute percentage errors (APE) were consistently below 10%, with the lowest APE of 5.30% in 2019. The highest coefficient of determination (R²) between the extracted and statistical data was 0.984 1, indicating an excellent fit and strong linear correlation. 3) The maize planting areas were 9 398.45, 11 487.37, 8 996.56, 7 690.33, 9 508.69, and 9 550.35 hectares, respectively, from 2019 to 2024 using the CBAM-UNet model. The temporal trend exhibited an initial increase from 2019 to 2020, followed by a subsequent decrease until 2022, and then a rebound with stabilization in the later years of 2023 and 2024. The research findings can provide technical support for the rapid identification and mapping of maize planting areas in the irrigation regions.