Evaluating changes with vegetation cover in PRISMA's spectral, spatial, and temporal attributes and their performance for classifying savannahs in Brazil

The recent advent of hyperspectral satellites with larger swath width than that of previous sampling missions brings new perspectives for mapping savannahs in Brazil. Here, we evaluated changes with vegetation cover in different spectral, spatial, and temporal attributes, derived from the PRecursore IperSpettrale della Missione Applicativa (PRISMA), and their performance for Random Forest (RF) classification of savannah physiognomies at the Brasilia National Park (BNP). To obtain the spectral attributes, we selected a PRISMA image acquired during the local dry season (August 17, 2020). We evaluated the classification performance of the reflectance of 166 bands, 22 vegetation indices (VIs), and four endmember fractions derived from a linear spectral mixture model (SMA). In addition, 24 parameters describing the depth, area, width, and asymmetry of the absorption bands centred at 680 nm (chlorophyll), 980 nm and 1200 nm (leaf water), and 1750 nm, 2100 nm and 2300 nm (lignin-cellulose) were also considered in the analysis. For the spatial attributes, we tested the performance of 8 Gray Level Co-occurrence Matrix (GLCM) metrics of image texture associated with the 864-nm near-infrared (NIR) band. In order to determine the temporal attributes, we considered other three PRISMA images obtained in 2020 (11 May, 4 September, and 3 October). Using these images, we calculated the rate of changes for each of the 22 VIs in the browning and greening periods of the savannah environment. A feature selection procedure was applied to the datasets. The results showed that the vegetation gradient from savannah grassland to woodland areas controlled the behavior of most attributes. For instance, the reflectance of the PRISMA NIR bands and the depth of the chlorophyll (680 nm) and leaf water (980 nm and 1200 nm) absorption bands increased with increasing vegetation cover. On the other hand, the reflectance of the visible and shortwave infrared (SWIR) bands and the depth of spectral features associated with non-photosynthetic vegetation followed the opposite pattern. Except for the metrics of image texture, the other spectral (reflectance, VIs, endmember fractions, and absorption band parameters) and temporal (browning and greening rates of vegetation changes) attributes had close classification performance before or after feature selection. When combined into a single dataset, gains of 15% in overall classification accuracy were observed when compared to the individual use of reflectance data in the analysis. From the seven savannah classes tested for classification, areas of woodland savannah, savannah grassland, and riparian forest were adequately mapped using this approach (F1-scores between 0.72 and 0.91).