Endmember spectral variability and index determination for retrieving fractional vegetation cover in the Loess Plateau

Accurate remote sensing retrieval of fractional vegetation cover (FVC) of photosynthetic vegetation (PV) and non-photosynthetic vegetation (NPV) is essential for assessing regional soil erosion. However, current linear spectral unmixing methods often ignore variability in endmember spectral indices, causing errors in FVC estimation. Using field-measured hyperspectral data and the derived indices of NDVI and the Cellulose Absorption Index (CAI), we analyzed the spectral properties of the endmembers across red, near-infrared, and shortwave infrared bands, examining their variability among different vegetation types and seasons. Furthermore, we identified optimal spectral indices and their combinations for retrieving PV, NPV, and bare soil (BS) fractions using MODIS imagery. Results showed that while endmember NDVI varied significantly with vegetation type (e.g., mean forest PV NDVI of 0.85 vs. 0.64 for grass) and season, the CAI demonstrated no significant variability under the same conditions. A three-component linear spectral unmixing model was developed and evaluated using MODIS-derived indices: NDVI, Enhanced VI (EVI), Kernel-NDVI (kNDVI), and two alternatives for CAI—Shortwave Infrared Ratio (SWIR32) and Dead Fuel Index (DFI). The kNDVI-SWIR32 and NDVI-SWIR32 combinations exhibited the highest predictive accuracy. Determination coefficients for FPV, FNPV, and FBS were 0.92, 0.74, and 0.70, respectively, with Nash-Sutcliffe efficiency coefficients of 0.90, 0.74, and 0.70, and RMSE values of 10.2 %, 16.6 %, and 13.5 %, respectively. This study provides a robust theoretical basis for high-precision retrieval of FPV and FNPV in the Loess Plateau and offers promising technical support for improving the accuracy of the cover and management factor in soil erosion models.