Low-Rank Hierarchical Clustering of PRISMA Hyperspectral Images to Identify Burned Areas

The analysis of post-fire satellite images is essential for assessing fire-affected areas. A common practice for identifying burned areas is the use of burn indices. Unfortunately, these indicators are selected just from some parts of the electromagnetic spectrum and then do not use all the information stored in the signal. In this work, we propose the use of a multivariate unsupervised approach for detecting burned areas from PRISMA hyperspectral images. Through hierarchical clustering and a two-rank nonnegative decomposition based on Nonnegative Matrix Factorization (NMF), we extract spectral signature information, known as endmembers, and their respective fractions within detected clusters. In particular, the use of NMF allows us to identify different burned areas based on their abundance variability. Spatial analysis was conducted to compare results from the proposed approach with both ground truth data and dNBR (differenced Normalized Burn Ratio) index. The experiments on different study areas from PRISMA mission show that the method can contribute to the understanding of the impact of wildfires on the environment.