SAM2-based multi-task learning for fuzzy land parcel boundary extraction in mountainous areas

【Objective】 The multi-task deep learning model of fuzzy boundary extraction was constructed to solve the problems of difficult extraction of fuzzy boundary and difficult elimination of pseudo-boundary, which could provide reference for the extraction of land parcel block boundary in areas with the broken land parcels. 【Method】 Taking Yizhou District, Hechi City, Guangxi as the research area, by interpreting the broken remote sensing image of a typical mountain area, the fuzzy boundary extraction data set was established, the SAM2 visual large model was introduced and the encoder was fine-optimized by using the Adapter, the auxiliary task of land attribute extraction was designed, and the multitask fuzzy boundary extraction deep learning model SAM2Xi was constructed. The results showed that the model was effective in extracting fuzzy boundary in mountainous area with broken land parcels. 【Result】 The SAM2Xi model performed the best on the global best threshold（ODS） and single graph best threshold（OIS）, which were 0.663 and 0.672 respectively, showing the highest edge detection accuracy and adaptability, but 50% accuracy recall（R50） were slightly lower than the DexiNed model. The SAM2Xi model combined semantic information and edge features to enhance the fuzzy boundary recognition ability, especially in complex scenes. SAM2Xi model still maintained high precision under low contrast and complex background, and the detail retention, coherence and noise suppression of fuzzy boundary region were better than other models. In addition, the SAM2Xi model performed the best in the pseudo-boundary clearing task, and its advanced feature extraction and optimization mechanism almost completely eliminated pseudo-boundary interference, maintaining high-precision edge detection in various scenarios, with higher robustness and accuracy. The SAM2Xi model could successfully extract the land parcel information of the study area（the total number of land parcel spots was 1587597, and the total area was 145696.646 ha）, and the extracted land parcel distribution was highly consistent with the actual situation, which was shown as follows:（1） it could accurately divide various land parcels within a large area of cultivated land;（2） a few pieces of cultivated land or garden could be extracted from the building;（3） it could extract the land parcels of forest land that could be divided separately（artificial forests）, but natural forests would not be misidentified. 【Conclusion】 The SAM2Xi model based on SAM2 multi-task learning realizes the double breakthrough of fuzzy boundary recognition and pseudo-boundary elimination, and has obvious advantages in complex terrain adaptability, boundary continuity maintenance and noise suppression, providing technical support for land parcel boundary extraction and precise management of agricultural resources in mountainous areas of southwestern China under complex terrain.