Ecological Informatics (Dec 2024)
Improved estimation of non-photosynthetic vegetation cover using a novel multispectral slope difference index with soil information, Sentinel-1 data, and machine learning
Abstract
Non-photosynthetic vegetation (NPV) plays a crucial role in vegetation–soil ecosystems by influencing the dynamic uptake of carbon, water, and nutrients. The accurate estimation of the fractional cover of NPV (FNPV) is vital for managing natural resources, modeling carbon dynamics, and monitoring vegetation systems. NPV indices (NPVIs) are widely utilized for estimating FNPV over large areas. However, distinguishing NPV from bare soil (BS) remains challenging owing to subtle differences in their multispectral reflectance and the interference of the soil background. To address this, we propose a novel NPV spectral slope difference index (NSSDI) based on the distinct spectral curve shapes of NPV and BS. The NSSDI incorporates three spectral slopes: two between the visible and near-infrared bands and one between the shortwave infrared 1 (SWIR1) and SWIR2 bands from Sentinel-2 (S2) and Landsat-8 (L8). Additionally, we investigate the integration of soil information and radar data to improve the FNPV estimation using three machine learning algorithms. Validation against in situ measurements reveals that the NSSDI from S2 is more sensitive to FNPV than the recently published NPV–soil separation index (NSSI), whereas the NSSDI from L8 outperforms traditional NPVIs. Incorporating soil properties such as soil organic matter and soil moisture improves the FNPV estimation accuracy compared with using S2 or L8 NPVIs alone. The combination of Sentinel-1 (S1) synthetic aperture radar (SAR) data with either optical satellite also enhances the retrieval accuracy. The Gaussian process regression (GPR) model, which integrates S2 NSSDI, soil data, and S1 SAR data, achieves an R2 of 0.78 and the lowest RMSE of 0.109. Similarly, the GPR model based L8 NPVIs, soil data, and SAR data attains an R2 of 0.71 and an RMSE of 0.128. Both GPR models outperform the random forest and XGBoost models. Monthly FNPV estimates from April to August 2019 in the Northern Territory, Australia demonstrates strong spatial consistency across both satellites using the GPR models in the Google Earth Engine. These results suggest that combining NPVIs with soil and SAR data can facilitate accurate large-scale FNPV estimations.
