Abstract: Illegal cultivation and reclamation activities are expressly prohibited in the area of reclamation of steep slopes. Large areas of sloping farmland have posed a high risk of soil erosion to reduce land productivity, even seriously threatening food security. However, manual ground surveys have been confined to investigating the illegal cultivation on the steep slopes and the prohibition of reclamation, due mainly to time-consuming and laborious. Meanwhile, the steep slopes are often distributed in dangerous areas, such as mountains and hills, leading to delays in the manual supervision of regional soil and water conservation. Fortunately, remote sensing interpretation and detection can be applied to monitor soil and water conservation in steep slopes, particularly with the rapid development of remote sensing and unmanned aerial vehicle (UAV) technology in recent years. Deep learning has been used to extract sloping farmland in the field of cultivated land identification. However, it is still lacking to consider the terrain slope, which has a great impact on soil erosion. At the same time, the large-scale remote sensing images often contain both concentrated contiguous fields and scattered small and micro fields; There are quite different directions and shapes of the fields, due to the differences in topography and landform, leading to the missed, false detection and low confidence in the detection of convolutional neural networks. The low detection efficiency and high time cost can also be caused by manual visual inspection in the supervision of illegal cultivation areas. In this study, a new model was designed to detect the remote sensing images from the illegal cultivation areas on steep slopes using the improved YOLOv9. The datasets were used from the GF-1 satellite and ALOS PALSAR. Firstly, 12.5 m resolution DEM was calculated to extract the potential slope farmland areas using ALOS data. Then the GF-1 satellite remote sensing images in the potential slope farmland area were utilized as experimental data. The specific procedures and main contribution were as follows :1) The lightweight self-attention mechanism (SimAM) was introduced into the backbone network of YOLOv9 for the richer background and the better discrimination of features; 2) The ultra-lightweight upsampling operator (DySample) was selected to replace the kernel-based dynamic one, in order to reduce the number of network parameters for the high recognition speed and accuracy; 3) The recognition accuracy of YOLO was verified for the illegal tillage areas after the introduction of SimAM and DySample; Finally, the comparison was made on the Faster RCNN, YOLOv5, YOLOv7, and Yolov8 models. Field investigation and remote-sensing image migration experiments were also carried out to verify the effectiveness of the improved model. The results show that the average recognition accuracy of the improved YOLOv9 model was 82.27%, which was better than that of mainstream object detection, such as Faster RCNN, YOLOv7, and YOLOv8. The high reliability and effectiveness were achieved in the improved model. The ablation test showed that the accuracy, recall, mean average accuracy, and F1 score of the improved model increased by 3.62, 3.78, 1.86, and 3.7 percentage points, respectively, compared with the original model. The findings can provide high-quality data support for soil and water conservation, particularly for the decision-making on the regional control of soil erosion, ecological environmental protection, as well as green and low-carbon sustainable production.