Land cover classification in high-resolution remote sensing imagery of natural and semi-natural scenes typically requires extensive manual annotation. Although the Segment Anything Model (SAM) exhibits powerful zero-shot segmentation capability, its class-agnostic nature and heavy reliance on manual prompts hinder fully automated semantic classification in complex natural scenes. This study introduces an adaptive enhanced Segment Anything Model framework (AE-SAM) that achieves fully automated land cover classification using only noisy, low-resolution global land cover products as priors. The framework incorporates a noise-resilient semantic alignment mechanism that reconstructs high-fidelity point prompts from low-resolution global land cover products through adaptive covariance-guided outlier filtering, multi-modal feature clustering, and spatial homogeneity constraints. These optimized prompts drive SAM to produce multiple hypothesis masks that are subsequently fused into a coherent classification map via adaptive optimization of model confidence and spatial consistency. Validated on 15 data sets with various resolutions (3 to 60 cm), AE-SAM achieved an average overall accuracy (OA) of 96. 52%, a Kappa coefficient of 0. 907, and a mean intersection over union (mIoU) of 0. 855. Notably, our method outperforms the zero-shot SAM Automatic Mask Generator by 7. 79 percentage points in OA, 0. 162 in Kappa coefficient, and 0. 138 in mIoU. It also surpasses few-shot segmentation approaches such as RSAM-Seg and traditional classifiers like Random Forest when all methods are evaluated under the few-shot setting. The results demonstrate that AE-SAM enables fully automated, high-accuracy land cover classification in natural and semi-natural environments using only coarse, noisy global priors—offering a practical solution for large-scale mapping in which manual annotation is infeasible. The source code is publicly available at https: //github. com/sophie1987/AE-SAM/blob/main/samᵤpdate. py.
Kui et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: