Towards fractional vegetation cover mapping in complex mountain environments through a knowledge-guided unmixing framework

Mountain vegetation dynamics serve as sensitive indicators of environmental change, making accurate fractional cover mapping of primary components crucial for understanding ecosystem responses. This task has been facilitated by multispectral satellite imagery, particularly Sentinel-2, which provides continuous observations. To translate these observations into fractional cover maps, existing studies typically apply fraction estimation models to synthetic spectra derived from field-measured signatures. However, limited field surveys in mountain terrain present a data bottleneck for synthetic spectra construction, while complex vegetation composition introduces spectral ambiguities that existing methods struggle to represent. Therefore, this study develops a knowledge-guided unmixing framework for annual mapping of the fractional cover of herbaceous vegetation, woody vegetation, and bare land in mountain environments using Sentinel-2 time series. Specifically, to overcome data bottleneck, we develop a knowledge-guided strategy that automatically builds adaptive physio-phenological feature spaces to extract representative signatures. To resolve spectral ambiguity, we incorporate nonlinear mixing mechanisms to synthesize spectra that better represent complex mixing relationships. The synthetic library was then used to train Unmix-Net, a spectral-temporal regression model for fraction estimation. The framework was applied to the central Qilian Mountains and southern Hengduan Mountains spanning broad climatic gradients, validated with high-resolution UAV data. It achieved MAEs of 12.38 %, 8.58 %, and 9.82 % for herbaceous, woody, and bare land, outperforming existing models by up to 7.80 %. When implemented across landscapes ranging from arid steppes and alpine meadows to tropical rainforests, the framework maintained comparable performance. The comparative experiments further confirmed that the proposed extraction strategy improved the balance of component estimates, while combining complementary mixing mechanisms reduced estimation errors. By integrating domain knowledge into data-driven learning, this framework supports spatially continuous fractional vegetation mapping in mountain environments.