Prevention and management of wildfires: vulnerability mapping and machine learning-based algorithm development for fuel mapping using hyperspectral imagery

Fire is a major ecological disturbance and threatening factor of ecosystem sustainability around the world and specifically in Mediterranean regions. Natural vegetation ecosystems are important environmental resources that provide various benefits to the human society whereas it also acts as fuel for wildfires. Hyperspectral imagery (HSI) is a passive technology which has the ability to classify the wildfire fuel types in a scene by means of several (hundreds) narrow band spectral acquisitions. This PhD thesis focused on developing a wildfire vulnerability map using GIS data for Sardinia and a procedure for wildfire fuel mapping using PRISMA HSI. Firstly, wildfire vulnerability map was generated using the vulnerability index comprising of the three main components: exposure, sensitivity and coping capacity. Exposure, representing the presence of assets (people, property and ecosystems) in areas where wildfires occur. Sensitivity, representing the degree to which these assets can be affected by a wildfire, linked to their predisposition to suffer certain type and magnitude of losses. Coping capacity, related to the measures applied to anticipate potential effects or to respond in case of fire occurs, based on institutional practices within several countries. Composite indices for each of the components were created using GIS data of population density, fuel types, location of protected areas, roads infrastructure and surveillance activities, taking into account the effect of the third dimension wherever is necessary. The additive type model was selected for the aggregation of components by allocating weights in the order of importance, mainly to differentiate the effects of individual elements and to streamline the interpretation of the outputs. Specifically, non-coping capacity was improved by including road density along with other institutional variables such as firefighters and surveillance areas. The vulnerability map of Sardinia developed by combining exposure map, sensitivity map and non-coping capacity map was shown. In this map, the value ranges from 0 to 1 representing from lower to higher vulnerable pixels correspondingly. Secondly, a semi-supervised machine learning approach for discriminating the wildfire fuel types was developed for the hyperspectral imagery (HSI) of PRISMA, a recently launched satellite of Italian Space Agency. Though machine learning classifiers provide better accuracy comparatively, many remote sensing specialists hesitate to use them because of the unavailability of required datasets. So, here, a procedure was developed to generate samples using single spectral signature as input data point for each class to apply support vector machine classifier and followed by, unmixing of mixed pixels by fully constrained linear mixing model. The procedure developed for classifying the fuel types available in the image of south-west Sardinia covering a part of Monte-Arcosu Forest and 18 different fuel types were classified in this region of interest. In order to correlate the classified fuel types to fuel models of Anderson or Scott/Burgan, further classification was carried out. Fuel types were classified according to the sparse/dense type, plain/mountainous type, open/closed type, and climatic conditions and for which available maps such as biomass, DEM, Tree Cover Density Map and iso-bioclimatic condition map were used respectively. Relative Greenness map was generated using time-series Sentinel-2 data. Then, the procedure of conversion from classified map to fuel map according to the JRC Anderson Codes and Scott/Burgan standard fuel models has been presented. The procedure was implemented on the HSI images obtained for south of Sardinian Island and for north-west of Latium in Italy as demonstration purpose. The classified map has been validated in different ways i.e. by using reference data, ground data and field data and obtained an overall accuracy of greater than 80% for all the cases. The stability of this approach was also tested by repeating the procedure on another HSI obtained on Latium in Italy and obtained degree of confidence greater than 95%. The proposed approach in this work can be used to generate wildfire fuel map using hyperspectral (PRISMA) data with higher accuracy over any part of Europe using LUCAS points as input. SWOT analysis has been conducted to understand the Strengths, Weaknesses, Opportunities and Threats of PRISMA hyperspectral imagery for wildfire fuel mapping. Though it is not possible to overcome all the weaknesses and threats, strategies to overcome some of them were discussed. Thus, the most vulnerable spots of wildfires can be referred using the developed vulnerability map whereas the wildfire fuels can be mapped-in for the areas of interest with hyperspectral image of PRISMA as per the proposed approach. Fuel map is useful to fire managers, researchers, policy makers and systems in applications such as study of fire behaviours, fire potential, fire emissions, carbon budget, fuel management, fire effects and ecosystem modelling. With this, it can be considered that this work has a major role in the prevention and management of wildfires.