Fire risk assessment

Abstract

The increasing demand for effective forest fire prevention instruments has faced operational and future Earth observation instruments with the challenge of producing updated and reliable maps of vegetation moisture. Various empirical band-ratio indexes have been proposed so far, based on multispectral remote sensing data, that have been found to be related to vegetation moisture expressed in terms of equivalent water thickness (EWT), which is defined as the weight of liquid water per unit leaf area. More sophisticated retrieval methodologies can be adopted when hyperspectral data are available, e.g. based on spectral curve fitting in selected water absorption bands or radiative transfer model inversion, allowing for better estimates of EWT. Problems arise with the evaluation of fuel moisture content (FMC), which is the percentage weight of water per unit of oven-dried leaf weight, due to its weak signal in vegetation spectrum. FMC is essential in fire models, and it is not interchangeable with EWT. Basing on simulated vegetation spectra, this study aims at demonstrating that hyperspectral images of vegetated areas can be effectively used to evaluate FMC with accuracies not achievable with multispectral data. To this purpose, radiative transfer models PROSPECT and SAILH have been used to simulate canopy reflectance. Vegetation spectra have then been convolved to hyperspectral data basing on the design specifications of a formerly planned ASI-CSA hyperspectral mission (JHM configuration C), similar to those of the forthcoming PRISMA. For comparison against multispectral instruments, measurements from the Operational Land Imager (OLI) have also been simulated. Two retrieval methods have been tested, based on spectral indexes and on partial least squares (PLS) regression. The latter methodology is particularly suited to analyse high-dimensional data. Results confirm that spectral indexes are good predictors of vegetation moisture expressed as EWT, but their performance in evaluating FMC is poor. By using PLS regression on hyperspectral data, a linear model can be built that accurately predicts FMC. No such result is achievable from OLI simulated data.