Characterization of forest biodiversity in Western Amazon using CAO-VSWIR imaging spectroscopy

Mapping canopy species richness is a key to the study and conservation of biological diversity in tropical forests, but to date, no reliable methods exist for operational biodiversity mapping of tropical regions. Airborne imaging spectroscopy has proven potential for the discrimination of canopy tree species, as a combination of high spectral and spatial resolution allows measurement of subtle spectral variations among individual tree crowns, corresponding to the chemical properties of the leaves in different species. We developed a method to estimate the Shannon diversity index, a popular biodiversity indicator, of a forest canopy from airborne spectral data by building upon the Spectral Variation Hypothesis, which relates biological diversity to spectral variability. We collected and analyzed hyperspectral data acquired by the Carnegie Airborne Observatory (CAO) Airborne Taxonomic Mapping System (AToMS) over the Los Amigos Conservation Concession in the Peruvian Amazon. The data have a spatial resolution of 2.0 m and 217 bands evenly spaced between 380 nm and 2510 nm. The method relies on a k-means clustering of a subset of pixels randomly selected from a site, each cluster serving as a proxy for different species. Each pixel in the image is then assigned to the nearest 'proxy-species', the Shannon index is computed for a given area, i.e. 1 ha, and the process is repeated several times to obtain the average estimated Shannon index. To test our approach, we applied the method to two types of data acquired by CAO AToMS. The first was an artificial gradient of biological diversity generated using pixels corresponding to species identified during a field campaign. This artificial gradient allowed total control on the number of species (ranging from 1 to 36 species per ha), and accurate quantification of the results. The spectral diversity index mapped using our method showed a strong correlation with the actual Shannon diversity index (R^2=0.81). The second dataset was a 2000 ha mapped area covering patches of primary and secondary forests and logging areas, resulting in a large range of Shannon index values. The assessed values on this second dataset also proved correlated with field measurements, and additional field measurements are currently being collected for validation. The method developed here is computationally efficient and reliable when processing large areas. Future directions include applications to watershed and regional scales, which will provide important inputs for tropical forest conservation and management.