Highlights What are the main findings? The optimized traditional non-photosynthetic vegetation cover inversion models were setup by incorporating spatial heterogeneity through covariance matrix integration, combined with spectral phenological weights. The optimized model implements spectral convolution to align hyperspectral endmembers with multispectral sensor characteristics, and the integration of spatial heterogeneity analysis has significantly improved the accuracy of non-photosynthetic vegetation detection, which has been implemented and validated in the arid regions of northwest China. What is the implication of the main finding? This study addresses the limitation where spectral discrepancies emerge between different regions during the identification of non-photosynthetic vegetation cover. The optimized non-photosynthetic vegetation identification model and spectral dataset enable dynamic long-term monitoring of non-photosynthetic vegetation, providing critical insights into the assessment of vegetation ecological health in arid and semi-arid regions under global warming.Highlights What are the main findings? The optimized traditional non-photosynthetic vegetation cover inversion models were setup by incorporating spatial heterogeneity through covariance matrix integration, combined with spectral phenological weights. The optimized model implements spectral convolution to align hyperspectral endmembers with multispectral sensor characteristics, and the integration of spatial heterogeneity analysis has significantly improved the accuracy of non-photosynthetic vegetation detection, which has been implemented and validated in the arid regions of northwest China. What is the implication of the main finding? This study addresses the limitation where spectral discrepancies emerge between different regions during the identification of non-photosynthetic vegetation cover. The optimized non-photosynthetic vegetation identification model and spectral dataset enable dynamic long-term monitoring of non-photosynthetic vegetation, providing critical insights into the assessment of vegetation ecological health in arid and semi-arid regions under global warming.Abstract Non-photosynthetic vegetation fractional cover (fNPV) is a key indicator of vegetation decline and ecological health. Traditional inversion models assume identical spectral signatures for the same vegetation cover class across entire study areas. Spectral variations occur among regions due to divergent soil properties and vegetation types. To address this limitation, extensive ground sampling was conducted; ground observation data from multiple regions were utilized to establish localized spectral libraries, thereby enhancing spectral variability representation within the study area while concurrently optimizing vegetation indices across different sensor systems. The results reveal that, within the optimized spectral mixture analysis model, the coefficient of determination (R2) for fNPV using the NPV soil separation index (NSSI) for Sentinel sensor is 0.6258, and that of fPV using the modified soil adjusted vegetation index (MSAVI) is 0.8055. The MSAVI-NSSI achieved an R2 of 0.7825 for fNPV and 0.8725 for photosynthetic vegetation fractional cover (fPV). Optimized vegetation indices also yielded favorable validation results. Landsat's theoretical predictions improved by 0.1725, with validated results up by 0.1635. MODIS showed improvements of 0.1365 and 0.1923, respectively. This enhancement significantly improves the accuracy of NPV fractional cover identification, providing critical insights for vegetation ecological health assessment in arid and semi-arid regions under global warming. Furthermore, by optimizing the spectral constraint weights in remote sensing images, a solution is provided for the long-term monitoring of vegetation health status.