Ecological Indicators (Oct 2024)
Enhancing chlorophyll content monitoring in coastal wetlands: Sentinel-2 and soil-removed semi-empirical models for phenotypically diverse Suaeda salsa
Abstract
In recent years, China has gradually begun restoring native salt marsh vegetation such as Suaeda salsa (S. salsa) in coastal wetlands that were damaged by the long-term invasion of Spartina alterniflora. Chlorophyll content (Cab), an important indicator of vegetation health, necessitates extensive and long-term monitoring using Sentinel-2. However, due to the influence of betacyanin (Beta), S. salsa exhibits different phenotypes (red and green) under various stress conditions, making remote sensing mechanism studies of this unique vegetation more challenging. In particular, satellite multispectral images are significantly affected by soil background in mixed pixels, making it imperative to mitigate this influence. This study explores the applicability of a recently proposed spectral separation of soil and vegetation (3SV) in Sentinel-2 multispectral and S. salsa vegetation from a remote sensing mechanism perspective, and further improves it. Additionally, a comparative analysis was conducted on the effectiveness of combining 3SV with several mainstream chlorophyll-sensitive indices. The advantages of machine learning algorithms were leveraged to develop a high-precision hybrid semi-empirical model for estimating Cab in different S. salsa phenotypes. The research findings indicate that: (1) The 3SV algorithm, adjusted with slope compensation and B2 and B4 bands, is applicable to green S. salsa scenarios. For red S. salsa scenarios, further adjustment using B2 and B3 bands and coverage fraction is required. (2) The MTCI, MRENDVI, MND, and MNDRE indices combined best with the modified 3SV, significantly reducing the RMSE of the semi-empirical models, especially under wet soil conditions with soil fraction fsoil < 0.5. (3) The highest accuracy (RMSE = 3.83 μg/cm2) for Cab estimation models for different S. salsa phenotypes was achieved by combining the modified 3SV soil-removed algorithm and the four indices with particle swarm optimization random forest regression (PSO-RFR).