High density biomass estimation: Testing the utility of Vegetation Indices and the Random Forest Regression algorithm

Accurate estimates of wetland above ground biomass (AGB) have increasingly been identified as a critical component for an efficient wetland monitoring and management system. Multispectral remote sensing based indices have proven inadequate in estimating biomass especially at high canopy density. In this study we investigated the use of vegetation indices derived from field hyperspectral data to estimate papyrus (Cyperus papyrus) biomass. Spectral and above ground biomass measurements were collected at three different areas in the Greater St Lucia Wetland Park, South Africa. We evaluated the potential of narrow-band normalized difference vegetation index (NDVI) calculated from all possible two band combinations between 700 nm to 1000 nm. Subsequently, we utilized random forest (RF) as a modeling tool for predicting papyrus biomass. The results showed that papyrus biomass can be estimated at full canopy level under swamp wetland conditions (R = 0.73, RMSEP = 276 g/m; 8.6 % of the mean). From our results, random forest has proved to be a robust feature selection method in identifying the minimum number (n = 4) of narrow-band NDVIs that offered the best overall predictive accuracy. The results can be scaled to spaceborne or airborne sensors such as Hyperion or HYMAP for predicting vegetation biomass in wetland areas using remotely sensed data.