Landscape pattern zoning-based classification for high-resolution orchard mapping in heterogeneous agricultural landscapes

High-resolution land-use classification in heterogeneous agricultural landscapes is challenged by geographically varying class–feature relationships that are difficult to represent with a single global classifier. We developed a Landscape Pattern Zoning-based Classification (LPZC) framework that transforms the composition and spatial configuration of historical land use into Landscape Pattern Zones (LPZs), which are then used to organize training samples and fit local classifiers. The framework was evaluated for orchard mapping from 2-m GF-6 imagery in Meishan, China, using fixed training, validation, and independent test partitions; LPZ parameters were selected on the validation set and locked before final testing. On the independent test set, LPZC increased F1-score from 0.557 for the global Random Forest baseline to 0.641 and increased recall from 0.526 to 0.675. A staged decomposition showed that this gain was driven primarily by zonal modelling rather than by zoning-guided sample allocation alone. LPZ performance depended jointly on thematic class representation, landscape-metric grain, and zoning extent, with validation F1-scores ranging from 0.381 to 0.576 across 45 candidate configurations. Perturbing historical land-use labels altered LPZ topology and reduced classification performance, but the response was non-monotonic with nominal perturbation level, showing that prior-data usefulness depends on whether informative spatial organization is preserved. In held-out comparisons, zone-count-matched LPZs achieved F1 advantages of approximately 0.005–0.019 over geomorphological and phenological zoning, while the locked LPZ exceeded four regular-grid schemes by 0.017–0.038 in F1-score. However, directly adding landscape-pattern metrics to a global classifier increased F1-score to 0.640, nearly matching LPZC while producing a different precision–recall balance. Additional Sentinel-2 features further increased LPZC F1-score to 0.661, indicating complementarity between landscape-based spatial priors and sensor-derived information. These results show that landscape information can support high-resolution classification through both predictor enrichment and modelling-domain restructuring. LPZC provides an explicit spatial framework for the latter by using historical landscape organisation to define locally structured modelling domains, although its geographic, temporal, and cross-class transferability requires further validation.