Fine-scale annual mapping and dynamic monitoring of shrub encroachment in alpine grasslands by integrating Sentinel-2 phenological features with Sentinel-1 structural information

Shrub encroachment can alter forage resources, soil–water processes, and landscape configuration in alpine grasslands, yet annual mapping remains challenging because shrub and herbaceous canopies converge spectrally and many apparent changes occur in diffuse ecotones. We developed a 10-m annual mapping framework integrating Sentinel-2 phenology, Sentinel-1 synthetic aperture radar (SAR) time-series information, and topographic covariates to distinguish shrubland from grassland in Maqu County on the northeastern Tibetan Plateau from 2018 to 2024. A learned SAR-derived shrubland-grassland discrimination score (SGDS) was constructed from annual Sentinel-1 descriptors using least absolute shrinkage and selection operator (LASSO)-regularized logistic regression to improve shrubland-grassland discrimination. In the 2024 benchmark experiment, the selected random forest (RF) configuration reduced classification error by approximately 17% relative to optical-only RF. SGDS outperformed the tested conventional SAR indices and the full SAR-statistics feature set, and adding SGDS increased RF accuracy from 91.26% to 92.47%. When the selected workflow was applied independently to each mapping year, annual overall accuracies ranged from 91.6% to 95.8%, based on the same 500 temporally screened reference locations spatially separated from the model-development samples. Nominal mapped shrubland area increased from 392.3 km2 in 2018 to 648.2 km2 in 2022 before declining to 560.4 km2 in 2024. Confidence and boundary-sensitivity analyses showed that the pronounced 2020–2021 increase was largely associated with low-confidence ecotonal pixels and decreased from + 96.91 to + 6.09 km2 as the mapped-class confidence threshold increased from 0.60 to 0.80, whereas the 2022–2024 decline remained negative under all tested confidence and boundary screens. Mapped shrubland area showed an apparent positive association with annual mean temperature, but this relationship was not robust after detrending and should be considered preliminary given the seven-year record. Overall, integrating learned SAR structural information with optical phenology improved annual shrubland-grassland mapping and provided a more reliable basis for interpreting mapped dynamics, with potential for locally calibrated application in comparable alpine regions.