Remote sensing-based prediction of forest fire characteristics

Forest fires are a major ecosystem disturbance at global scale, put pressure on agencies in charge of citizens and infrastructure security and cause unvaluable human losses. Fires are controlled by multiple static and dynamic drivers related to topography, land cover, climate, weather, and anthropic activity. Among these, weather is an active driver of live and dead fuel moisture, which has a direct effect on fire occurrence and behaviour. As a result, in areas experiencing prolonged droughts and heat waves, altered meteorological patterns lead to increased frequency and intensity of forest fires. The operational response of governments, local authorities, forest managers and civil protection agencies in charge of managing forest fires is informed by the assessment of factors controlling fire occurrence and behaviour, often synthesised in maps of fire danger. Danger is defined as the resultant of all factors affecting the inception, spread, and difficulty of control of fires, and it is typically expressed in the form of an index. Key contributors to fire danger are fuel type, amount, and conditions, notably with respect to moisture content. Remote sensing measurements in the shortwave infrared are sensitive to water content of live fuels, while measurements in the thermal infrared allow the detection of vegetation stress conditions due to vapour pressure deficit. In fact, several scholars proved that satellite estimates of vegetation water content and of land surface temperature could be effectively used to predict fire occurrence. Nevertheless, to the best of this author’s knowledge, no research was previously published connecting pre-fire remote sensing measurements to fire behaviour characteristics. This clearly identifies a knowledge gap which needs further investigation and that can be translated in the following research question: to what extent can remote sensing of forest condition be used to predict fire behaviour characteristics and assess the probability of extreme events? The research described in this dissertation aimed at developing methods based on pre-fire optical and thermal remote sensing observations of forests for the prediction of fire behaviour characteristics. The study was carried out in Campania, Italy (13595 km2), one of the most densely populated and fire affected regions in the Mediterranean. Data on all fire events recorded between 2002 and 2011 was provided by Carabinieri (Italian national gendarmerie) forest fire preparedness unit (Nucleo Informativo Antincendio Boschivo, NIAB). The study made use of MODIS land surface temperature (LST) and surface reflectance collection 6 products, which are publicly available on the USGS Land Processes Distributed Active Archive Center (LP DAAC). Approach was probabilistic in nature, trying to relate pre-fire satellite observations of vegetation conditions to the probability distributions of burned area, fire duration and rate of spread. Efforts initially focussed on assessing LST anomaly and its effect on fire behaviour characteristics. LST anomaly is a measure of excess enthalpy stored in fuels. It controls the probability of flames extinction and thus fire duration. First, a climatology of LST was constructed from the longest available time series of daily MODIS LST by means of the Harmonic Analysis of Time Series (HANTS) algorithm. HANTS was then used to construct annual models of daily LST. Finally, the daily LST anomaly was evaluated as the difference between the annual model and the climatology. Fires in the database were then associated with LST anomaly values recorded at their corresponding location on the day prior to the event. Probability distribution functions of log-transformed burned area (normal), log-transformed fire duration (generalised extreme value, GEV) and log-transformed rate of spread (Weibull) where then determined in ten decile bins of LST anomaly. The mean and the standard deviation of the normal distribution of log-transformed burned area showed a clear linear dependence on LST anomaly (r2=0.81, p<0.001 and r2=0.52, p<0.05 respectively), indicating an increase in the probability of large fires with increasing LST anomaly. Similarly, a marked linear dependence on LST anomaly was found for the location (r2=0.78, p<0.001), scale (r2=0.79, p<0.001) and shape (r2=0.87, p<0.001) of the GEV distribution of log-transformed fire duration, favouring longer fire duration with increasing LST anomaly. Conversely, the LST anomaly had a limited effect on the Weibull distribution of log-transformed rate of spread, with scale and shape showing slightly decreasing trends (r2=0.50, p<0.05 and r2=0.54, p<0.05 respectively). A likelihood ratio test showed that the probability models of log-transformed burned area, fire duration and rate of spread conditional to LST anomaly (alternative models) allowed the rejection of the corresponding unconditional models fitting all data (null models), confirming that LST anomaly is a covariate of burned area, fire duration and, to a lesser extent, rate of spread. These results are in line with expectations from models of the combustion process. Following a similar line of reasoning, this study further focussed on remote sensing of live fuel moisture content (LFMC). This vegetation property controls ignition delay, and thus affects flames propagation. The first step was the construction of a novel spectral index, the perpendicular moisture index (PMI), specifically designed to be sensitive to LFMC. The PMI was developed from simulated vegetation spectral data convolved to MODIS bands by noting that in the spectral reflectance subspace of MODIS bands 2 (0.86 µm) and 5 (1.24 µm) isolines of LFMC can be identified, and that these isolines are straight and parallel. By taking as a reference the line corresponding to LFMC=0 (completely dry vegetation), the PMI was calculated as the distance of measured reflectance from the reference line. The PMI is thus a measure of LFMC, and higher values of PMI correspond to higher moisture content. The index was found to be linearly related to LFMC, especially for dense vegetation cover (r2=0.70 when leaf area index is larger than 2, r2=0.87 when larger than 4). When vegetation cover is less dense, the contribution of soil background to the measured reflectance increases, and the PMI underestimates LFMC. PMI maps were produced from the MODIS 8-day composited reflectance product, and fires in the database were associated with the corresponding PMI value at the fire location in the pre-fire compositing period. Using the same approach adopted for LST anomaly, the probability distribution functions of log-transformed burned area, fire duration and rate of spread were determined in ten decile bins of PMI. The mean of the normal distribution of log-transformed burned area showed a clear linear dependence on PMI (r2=0.80, p<0.001), while no trend could be observed for standard deviation. A clear linear dependence on PMI was also found for scale and shape of the Weibull distribution of log-transformed rate of spread (r2=0.97, p<0.001 and r2=0.82, p<0.001 respectively). These results were further confirmed by a likelihood ratio test where the probability models of log-transformed burned area and rate of spread conditional to PMI allowed the rejection of the corresponding unconditional models fitting all data. Location and shape of the GEV distribution of log-transformed fire duration showed no significant linear trend with PMI, whereas scale showed a weak trend (r2=0.55, p<0.05). However, in the likelihood ratio test the probability model of log-transformed fire duration conditional to PMI failed to reject the corresponding unconditional model. These results showed that PMI is a covariate of burned area and rate of spread, as expected from flames propagation models, but not of fire duration. Predictions of fire characteristics based on concurrent observations of LST anomaly and PMI were compared with predictions based on the Fire Weather Index (FWI) System. This fire danger rating tool proved to be effective in several areas worldwide, including Europe. FWI values from weather reanalysis data were associated with fires in the database and were analysed with the same approach adopted for LST anomaly and PMI. It was found that parameters of the probability distribution function of log-transformed burned area and fire duration conditional to FWI System components followed clear linear trends, with increasing danger values leading to higher probabilities of large burned areas and long fire durations. Conversely, FWI System components were unrelated to the rate of spread. Trend analysis (coefficient of determination and p-value of the linear fit, Sen’s slope and Mann-Kendall test) and likelihood ratio tests were used to compare the trends in the parameters of the probability distributions of fire characteristics. It was shown that remote sensing predictions of burned area and fire duration were comparable or better than those from FWI, and that PMI is a good predictor of the rate of spread whereas FWI System components are not. The identified linear trends in the dependence of the parameters of the probability distribution of log-transformed burned area, fire duration and rate of spread on LST anomaly and on PMI allow the prediction of the probability of extreme events, conditional to ignition, as a function of pre-fire remote sensing observations. As both LST anomaly and PMI are good covariates of burned area, these two remote sensing observations of vegetation conditions can be used jointly to improve the prediction of the probability of fires larger than say, the 95th percentile of all events recorded in the study area (30 ha). It was found that the probability of a fire resulting in a burned area larger than 30 ha increases from 0.9% to 9.2% with pre-fire LST anomaly increasing from -2.1 to 4.3 K and increases from 1.8% to 7.4% with pre-fire PMI decreasing from 0.052 to -0.032. When the probability of fires exceeding 30.0 ha is modelled as a function of both LST anomaly and PMI, the probability increases from 0.5% to 12.7%. This confirms that the joint use of LST anomaly and PMI leads to improved predictions. The scientific community showed a consensus on the need to improve fire danger prediction through a more accurate assessment of live fuel condition. Existing fire danger rating systems estimate fuel moisture content from meteorological variables, which results in an undesired approximated solution due to underlying assumptions. Consequently, any direct observation of fuel moisture content has the potential to enable a better evaluation of fire occurrence and fire danger indices. From a remote sensing perspective, these considerations are translated in the research question on the need to understand to what extent can satellite measurements be used to predict forest fire behaviour characteristics. This research showed that remote sensing of vegetation in the optical and thermal domains allows the prediction of the probability distributions of fire behaviour characteristics such as burned area, duration, and rate of spread. These can be further used to evaluate the probability of extreme events, conditional to ignition, as a function of pre-fire remote sensing measurements, contributing to predict danger. It should be noted once more that this result was achieved by using pre-fire remote sensing observations, allowing the prediction of fire characteristics. In perspective, results showed in this dissertation can support the development of operational tools for forest managers and civil protection agencies in their fire preparedness activities.