Rapid and accurate acquisition of cultivated land change information is of great significance to food security management. This paper aims at the problem that remote sensing semantic segmentation method has many errors and omissions due to insufficient model applicability in large-scale and high-resolution image cultivated land non-agricultural detection. Multiscale Scene Classification-Xception （MSC-Xception）， a multi-scale scene classification method for high-resolution cultivated land images based on Xception， is proposed. The convolutional attention module CBAM is embedded into the output layer of the lightweight scene classification network Xception， which has outstanding performance in cultivated land scene classification， to enhance the model's ability to extract channel and spatial features. At the same time， the problem of low separation degree and rough details of mixed scenes existing in the single-scale scene-level classification in large-scale cultivated land extraction is also overcome. Firstly， a feature fusion method of multi-scale cultivated land scene is introduced to improve the separation degree of mixed scene， and then the boundary constraint of multi-scale segmentation vector is used to achieve the boundary refinement of scene-level classification. Compared with the typical Unet， PSPNet and DeeplabV3+ semantic segmentation methods， this method can better reduce the missed detection of large map spots， and the recall rate and F1 score in the cultivated land extraction experiment of GF-2 images in Qixia District in April 2018 increased by at least 15.1 percentage points and 8.8 percentage points respectively. In the non-agricultural detection of cultivated land in Qixia District from 2018 to 2022， the recall rate of suspicious spots increased by at least 7.16 percentage points.