Reframing historical sample reuse under forest succession: A geography-aware transferability framework for mountainous forest-type mapping

Forest-type mapping in mountainous regions is fundamentally constrained by the reuse of historical forest inventory samples whose spectral appearance may remain stable while their ecological and semantic meanings drift under forest succession. Conventional mapping strategies often assume temporal stationarity of training samples or attempt to mitigate inconsistencies through model retraining, without explicitly evaluating whether historical samples remain ecologically and geographically transferable to target-year conditions. To address this issues, this study reframs historical sample reuse as problem of transferability and proposes a forest-succession-aware sample transfer (FSAST) framework to explicitly constain sample reuse. FSAST integrates bi-temporal spectral stability, long-term sucdessional trajectory consistrncy, and multidimensional environmental similarity to identify historical samples whose ecological states remain representative of target-year forest conditions. The framework was implemented using Sentinel-1, Sentinel-2, and Landsat data within a standard random forest classification scheme and evaluated in Yunnan Province, China, a representative mountainous region with pronounced elevation-driven environmental gradients and diverse forest types. Compared with direct historical sample reuse (overall accuracy, OA = 69.94%) and a change-detection-based transfer strategy (OA = 73.27%), FSAST achieved a substantially higher classification accuracy (OA = 83.33%), with particularly pronounced improvements for forest types sensitive to successional dynamics, such as mixed and bamboo forests. These results demonstrate that classification reliability in mountainous environments is governed less by the quantity of training samples than by their ecological validity and geographical coherence. By explicitly accounting for forest succession, FSAST provides a data-centric and generalizable framework for responsible reuse of legacy forest inventory samples in environmentally heterogeneous regions.