Unmixing halophytes in coastal wetlands using a multi-sensor approach at two spatial scales

Coastal wetlands are geomorphic systems highly sensitive to shifts in environmental forcings such as variations in fluvial sediment transport rates, sea level rise, subsidence rates, nutrient concentrations, temperature, and atmospheric CO₂ levels. Despite these influences, the presence of vegetation growing on salt marshes significantly enhances their resilience. It mitigates surface and lateral erosion while fostering the accumulation of organic matter, which contributes to marsh soil accretion and the sequestration of organic carbon. Therefore, the characterization of vegetation properties, canopy biomass and species distribution, is crucial to provide a quantitative basis for bio-geomorphic modeling in coastal wetlands.

This study aims to spatially characterize key parameters—such as vegetation species distribution and biomass production—through repeated observations utilizing drone, airborne, and satellite multispectral (MS) imaging. The chosen site for this investigation is North Inlet in South Carolina (USA), renowned for its extensive tidal marshes supporting diverse vegetation species. MS and field data acquisitions were conducted in summer (August 2022 and August 2023), coinciding with the period of maximum biomass, and in winter (February 2023), corresponding to the phase of lowest biomass.

The application of a random forest (RF) approach proved highly effective in the unmixing process of halophytic vegetation species, enabling the retrieval of the percentage cover for each species. To train the algorithm, field observations were employed to classify drone-captured images within a limited section of the marshland. The random forest classification (RFC) algorithm achieves high accuracies in the classification of vegetation species based on the drone image, with a spatial resolution of about 0.02m and the overall accuracy of about 0.99. Based on this classification result, we applied the random forest regression (RFR) algorithm to unmix vegetation species using coarser-resolution WorldView2 data (pansharpened data with pixel of 0.5 × 0.5 m). Our results suggest that RFR achieves high accuracy in the unmixing process (0.80<R2<0.96 and 0.06<RMSE<0.14), enabling us to map the percentage cover of each species or bare soil over the entire North Inlet area. Furthermore, our field observations in August 2023 indicate strong correlations between Vegetation Indexes (VIs) derived from MS data, such as Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), Chlorophyll Index Green (CIg), and Chlorophyll Index Red (CIr), and marsh above-ground biomass (AGB). This suggests the potential utility of using multi-sensor MS data at various spatial scales to estimate marsh AGB.

We plan to incorporate field data from summer 2022 and winter 2023 to enhance the relationship between VIs and AGB, facilitating the estimation of AGB distribution over marshes. This analysis will be crucial for informing a spatially explicit bio-geomorphic model for marsh evolution (National Science Foundation Grant No. 2016068, project title: "Coupled Ecological-Geomorphological Response of Coastal Wetlands to Environmental Change").