Abstract: Fine-scale classification data of forests and grasslands are essential for ecological protection, natural resource management, and environmental monitoring. In extensive mountainous regions, strong spectral mixing between forests and grasslands complicates the accurate delineation of their boundaries, while the lack of effective characteristics limits the refinement of forest type classification. This study proposes a fine classification method utilizing Landsat TM/ETM + /OLI data (1991–2021) to construct long-term temporal characteristics. A characteristic set of NDVI time series (NDVI standard curve, NM, and NP) was developed to delineate forests and grasslands and identify their transition (forest–steppe ecotone) based on NDVI differences. Meanwhile, seasonal metrics—seasonal variation characteristics of canopy closure were then extracted to classify 5 forest types, and the distribution above the tree line was used to identify 2 grassland types. Distinct from forests or grasslands, 6 additional land cover types were mapped using Maximum Likelihood Classification (MLC) based on spectral characteristics, producing an integrated 14-class land cover map. The final classification achieved User’s Accuracy (UA) values of 94.85%, 87.80%, and 84.72% for forests, grasslands, and the forest–steppe ecotone, respectively. Compared with classifications using only NDVI standard curve or spectral characteristics, the proposed NDVI characteristic set improved accuracy by 4.93 and 11.92% for forests, 7.50 and 12.43% for grasslands, and 12.98 and 34.52% for the ecotone. All forest and grassland subtypes exceeded 80% accuracy. Compared with existing datasets, such as GlobeLand30, CLCD, and FROM-GLC (2017), the method more precisely captured the forest-steppe ecotone and enhanced the refinement of forest and grassland classification, improving forest classification accuracy by 8.07%, 6.87%, and 8.77%, and grassland accuracy by 7.36%, 25.96%, and 12.66%, respectively. The current study establishes a basis for investigating the physical attributes of the land cover classification model. This study offers novel concepts for developing efficient remote sensing classification characteristics for land cover.

Key words: Multi-temporal classification, NDVI time series, Landsat, Canopy closure, Tree line extraction