Evaluating automated endmember extraction for classifying hyperspectral data and deriving spectral parameters for monitoring forest vegetation health.

Abstract

The hyperspectral remote sensing datasets possess high capability in differentiating the spectrally similar features, and thus, they are immensely important in various forestry activities, especially vegetation classifications. But extracting endmembers for data training is a challenging task. The present study is focused on the use of automated endmember extraction technique for deriving endmembers during the unavailability of ground spectra. We used the Sequential Maximum Angle Convex Cone (SMACC) method on EO-1 Hyperion data for endmember extraction in the Barkot forest range of Dehradun district, Uttarakhand which were used for classification of the study area using support vector machine (SVM). Further, we estimated the vegetation health of the region by assigning the threshold weights for various derived environmental variables such as NDVI (Normalised Difference Vegetation Index), CRI (Carotenoid Reflectance Index), Anthocyanin Reflectance Index (ARI), Modified Simple Ratio (MSR), Modified Chlorophyll Absorption Ratio Index (MCARI) and WBI (Water Band Index). Then, to further validate the health of the forest types, we correlated it with the Land Surface Temperature (LST) from LANDSAT 5 ETM + data. The results showed a high classification accuracy of 89.13%. The healthy vegetation area coverage of the area was about 78.6% with most healthy class as Tectona grandis and Shorea robusta and its correlation with LST showed lower temperature range in healthy vegetation areas and vice versa. The study was useful in determining the superiority of SMACC automated endmember extraction and estimating the vegetation health.