Abstract: Combining semantic segmentation and edge detection deep learning networks, a Psi-Net deep learning network was constructed. The results showed that the Psi-Net deep learning network could effectively identify cultivated land, and the vast majority of cultivated land was effectively extracted with an Accuracy of 96.3%, a production accuracy （PA） of 98.1%, and a user accuracy （UA） of 97.1%. The Psi-Net deep learning network had limited the scope of cultivated land for boundary recognition, reducing the recognition of non-cultivated land boundaries. The completeness was 74.3%, the correctness was 80.2%, and the quality was 62.8%. The Psi-Net deep learning network could effectively identify the range of surface cultivated land, and under the limitation of cultivated land range, the extracted plot scale cultivated land boundaries all fell within the cultivated land range, without recognizing the plot boundaries outside the cultivated land, effectively reducing the impact of noise in extracting plot boundaries.

Key words: Psi-Net deep learning network, high spatial resolution, cultivated land plots, remote sensing image