Salt marsh elevation and habitat mapping using hyperspectral and LIDAR data

- A decision tree was used to combine salt marsh hyperspectral and elevation data.
- Nine salt marsh cover classes were mapped with a 90% overall accuracy.
- The decision tree reduced the confusion between Spartina alterniflora and mud pixels.
- The classification was combined with a LIDAR-derived DEM to correct elevation errors.
- Class-specific correction factors reduced DEM overall mean error and RMSE.

Accurate mapping of both elevation and plant distributions in salt marshes is important for management and conservation goals. Although light detection and ranging (LIDAR) is effective at measuring surface elevations, laser penetration is limited in dense salt marsh vegetation. In a previous study, we found that LIDAR-derived digital elevation model (DEM) error varied with vegetation cover. We derived cover-class-specific correction factors to reduce these errors, including separate corrections for three different height classes of Spartina alterniflora, the dominant macrophyte in southeastern U.S. salt marshes. In order to apply these cover class-specific corrections, it is necessary to have information on the distribution of cover classes in a LIDAR-derived DEM. Hyperspectral imagery has been shown to be suitable for the separation of salt marsh vegetation species by spectral signatures, and can be used to determine cover classes; however, there is persistent confusion both among the different height classes of S. alterniflora and between plants and mud (the Spartina problem). This paper presents a method to overcome the respective limitations of LIDAR and hyperspectral imagery through the use of multisensor data. An initial classification of hyperspectral imagery based on the maximum likelihood classification algorithm was used in a decision tree in combination with elevation and normalized difference vegetation index (NDVI) derived from the hyperspectral imagery to map nine salt marsh cover classes. The decision tree appreciably reduced the Spartina problem by reassigning classes using these ancillary data and resulted in a final overall classification accuracy of 90%, with a quantity disagreement of 1% and an allocation disagreement of 9%. The resulting hyperspectral image classification was then used as the basis for applying cover class-specific elevation correction factors to the LIDAR-derived DEM. Applying these correction factors greatly improved the accuracy of the DEM: overall mean error decreased from 0.10 ± 0.12 (SD) to − 0.003 ± 0.10 m, and root mean squared error from 0.15 to 0.10 m. Our results suggest that the use of decision trees to combine elevation and spectral information can aid both hyperspectral image classification and DEM elevation mapping.