Estimation of vegetation parameters by means of hyperspectral data and neural networks

Accurate knowledge of the distribution of vegetation can form a critical component for managing ecosystems and preserving biological diversity. Satellite multi-spectral remote sensing images have already been involved in vegetation types classification. The multi-spectral sensors include the Advanced Very High Resolution Radiometer (AVHRR), the Landsat Thematic Mapper (TM), and SPOT HRV. However, multi-spectral images have poor capability of discriminating forest species precisely. With the development of imaging spectroscopy, it has been found that the vegetation communities can be better differentiated using their hyperspectral reflectance in the visible to shortwave infrared spectral range. On the other hand, application of hyperspectral images also brings some problems: high-dimensional datasets have extremely huge volume and the narrow band tends to be strongly related with the adjacent ones, hence even useful information is immersed in useless signals. Therefore feature extraction is necessary for hyperspectral classification. Conventional Principal Component Analysis (PCA) is one of the most commonly used feature extraction techniques, which is based on extracting the axes where the data shows the highest variability, but the results of PCA maybe not related strongly with the characteristics of land cover class. In this work we present a methodology based on neural networks for extracting nonlinear principal components (NLPC) from hyperspectral data in vegetation analysis (forest and grassland). In the first stage of the study, the NLPC of the vegetation types were analysed looking for differences between type classes. In a second stage an inversion scheme, always based on neural networks, has been designed for the production of a classification map. The area located in alpine region (South Tyrol-Northern Italy) was the selected test site for the study. Indeed, hyperspectral data provided by the Hyperion and CHRIS platforms and concurrent extended ground-truth were available for this area. Further investigations also focused on possible synergies between hyperspectral data and multi-spectral and/or SAR data.