In mountainous and hilly areas, cultivated land resources are scarce, and the phenomenon of cultivated land fragmentation is serious. The long and narrow arable land plots are complex in structure, which makes it difficult to obtain the information of cultivated land at the land level quickly and accurately, and hinders the application of precise digital agricultural services based on high resolution remote sensing images in mountainous and hilly areas. Parcle-level cultivated land information can intuitively show the spatial distribution, boundary details of farmland, and is of great significance for precision agriculture management, distribution of planting subsidies, and agricultural resource survey. Existing edge detection/semantic segmentation networks-based farmland extraction methods ignore the structural features of the parcel, thus have limit performance for handling narrow and small plots, and there is also a blurring boundary problem. To address these issues, we proposed an accurate extraction method of cropland in mountainous area based on geographic parcels. This method combines the advantages of semantic segmentation and edge detection, and effectively extracts and integrates the linear features of the boundary and the internal texture features of the parcel, so as to improve the recognition accuracy of the cultivated land. The main features of the model in this paper are as follows: 1) The edge of cultivated land is regarded as a new class independent of cultivated land parcels, so that the semantic segmentation network can better distinguish the edge and internal area of cultivated land parcels; 2) A cascaded semantic segmentation and edge detection network is introduced to correlate the prediction of cultivated land surface and line, realize the fusion of boundary and texture features of cultivated land parcels and strengthen the edge features of cultivated land, so as to improving the accuracy of cultivated land block edge detection; 3) A focus training technique is proposed to address the problem that the edge pixels of cultivated land are far fewer than non-edge pixels, by enforcing the model pay more attention to the important but underrepresented edge pixels in high resolution remote sensing images in the training process, so as to improve the edge detection accuracy. We conduct experiments in Shaodong County, Hunan Province in the southwest mountainous area, using the Google Earth high-resolution remote sensing images as the data source, with a spatial resolution of 0.53 m. After manual selection, a total of 1000 512×512 image patches are obtained. Among them, 600 pieces are used as the training set, 200 pieces are used as the verification set, and 200 pieces are used as the test set. Experimental results show that the presented model achieves satisfying results with an overall accuracy of 92.91% and IoU (Intersection-Over-Union) of 82.84% on the test set, which was 4.28 percentage points and 8.01 percentage points higher than the baseline method respectively. Compared with the existing methods, the cultivated land extracted in this study is more consistent with the actual distribution pattern of cultivated land, which provides a practical method for accurate extraction of cultivated land information at the plot scale in mountainous and hilly regions.