Abstract
Greenhouses are vital for food security and agricultural modernization, yet their classification in remote sensing imagery is challenging due to scattered distribution, small scale, and spectral similarities. This study proposes a remote sensing classification framework using sample migration and dynamic threshold optimization for accurate, scalable greenhouse detection. Historical samples are migrated to the target area based on spectral similarity, reducing reliance on new labeled datasets and improving cross-regional generalization. A post-classification threshold optimization module corrects misclassifications using multi-dimensional indices, enhancing robustness in complex spectral environments. Empirical validation across six provinces in southern China showed superior performance (96.48% OA, 94.36% Kappa), outperforming traditional methods. This framework enables precise mapping of small-scale greenhouses in heterogeneous regions, supporting agricultural monitoring. It aids government agencies in crop area estimation, land-use tracking, and precision agriculture, while private enterprises benefit for asset evaluation and planning, promoting sustainable resource management and addressing large-scale land cover classification challenges.