Applications of Hyper-spectral and Radar Remote Sensing analysis: a case study of forest landscapes in Costa Rica

Abstract

The main objective this research was to assess two types of emerging remote sensing technology, hyper-spectral and SAR sensors, for an exploratory data analysis of land covers in the south of Costa Rica. Hyper-spectral data contain information in several narrow spectral bands in the optical domain, which give information on the biochemical and structural properties of vegetation, while the SAR data, as an active system, can penetrate the clouds making it a promising tool for ecosystem monitoring. The main hypothesis was that these two datasets would permit greater understanding of the spectral confusion between different land covers. From the hyper-spectral point of view, this knowledge could help to select and derive spectral signatures which serve as training data sets in supervised species classification in the optical domain. In the microwave domain, fusion and derived bands increase the separability and permit greater forest/non-forest classification accuracy in non-flat terrains. The hyper-spectral information is based on two information sources. The first comes from two scenes of the space-borne Earth Observing-1 mission with the Hyperion sensor and two scenes of the airborne hyper-spectral sensor HyMap. The second hyper-spectral data source was acquired from the field-based hyper-spectral clip-prove system. Furthermore, the microwave information corresponds to the TerraSAR-X HH and VV polarized images. Working in different land covers including Gmelina arborea plantations in the south of Costa Rica, individual Regions of Interest were manually digitized with reference to high spatial resolution aerial photographs datasets. Principal components of hyper-spectral space and airborne data were derived to perform a classification using two different approaches of Hierarchical Cluster Analysis. Spectra from field-based hyper-spectral clip-prove data was acquired from Gmelina arborea leaves in three plantations of 6, 8 and 18 years. Other reference spectra of land covers were also measured. With seven TerraSAR-X polarized images, principal component analysis as fusion technique and derived bands ratios were generated in order to evaluate the availability of reducing the speckle noise in non flat terrains to classify forest in the south of Costa Rica. The highest scene based spectra variability was in the Near Infra-Red portion of the electromagnetic spectrum. Hierarchical Cluster Analysis applied to the hyper-spectral scenes showed that cluster solutions of the PCs spectra from the two sensors present different separability solutions. The clusters solutions were subject to systematic differences; only one scene of EO-1 Hyperion and one of HyMap PCs spectra did not present spectral confusion among Gmelina arborea, palm oil and the forest. That indicates that the same sensor under different conditions will give different spectra and different cluster results. These results suggested that hyper-spectral imagery need not to be acquired at a very high spatial resolution to provide adequate discrimination of land covers. Furthermore, spectra collection and analysis are needed to acquire time series spectral signatures. The best Hierarchical Cluster Analysis classification was with the Approximately Unbiased p-values which permit the identification of clusters that exist at a predefined level of significance. Canopy phenology, a property related to the different acquisition times and atmospheric conditions, was important in clustering land covers. Regarding the field based spectra, there was spectral confusion in the majority of 18 years of leaves of Gmelina arborea and mangrove. Also, 6 spectra of this age were not clustered at all. There was spectral confusion between the spectra of Gmelina arborea leaves of 6, 8 and 18 years. However, the reflectance of field based spectrometers should be interpreted with caution. Sampling is a key factor as well as a challenge in leaf spectral analysis. Hyper-spectral and Synthetic Aperture Radar data was useful for land cover discrimination, and it did provide an unprecedented potential to classify forest and non-forest in tropical environments and avoid spectral confusion with highly related land covers. However, all the associated variability of acquisition parameters has be to taken into account in order to provide acceptable levels of accuracy.