Spatiotemporal analysis of vegetation fires using satellite data.

Abstract

Although vegetation fires are key in maintaining the savanna ecosystem, their uncontrolled occurrence profoundly threatens ecosystem stability, economies, and human safety. The increased risk of climate change requires robust spatiotemporal analysis methods to understand the impact of fire on ecosystems. Additionally, the accurate prediction of vegetation fire and the associated key drivers are critical in understanding fire regimes and the implementation of effective fire management strategies. This research utilised Moderate Resolution Imaging Spectroradiometer (MODIS) satellite data to analyse the spatiotemporal dynamics of vegetation fires. The first objective focused on systematically reviewing literature on the effects of burning on various ecosystem services. The reviewed articles were extracted from Elsevier’s Scopus, Web of Science and PubMed databases and analysed based on the Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) method. The findings from the review highlighted that there has been an increase in publications since 2010 and most studies were carried out in Asia and the United States of America. The most common satellite data used for analysing the effects of burning on ecosystem services was Landsat, whilst information on fire occurrence was extracted from the MODIS satellite data. Very few studies utilised AVIRIS, PlanetScope, and ASTER satellite data. Moreso, findings from the review revealed fire as a threat to grassland, forest, soil and wetland ecosystems with the forest landscapes being widely studied. The atmosphere is also affected by vegetation fires through particulate matter and carbon emissions. The second objective focused on detecting fire intensity hotspots and cold spots in Zimbabwe by utilising spatial statistics and MODIS-derived fire radiative power (FRP), a proxy for fire intensity. The variability of fire intensity clusters within various topographic and vegetation conditions was also analysed. The results indicated that most (44%) of the vegetation fires remotely sensed in Zimbabwe by the MODIS satellite sensor were of low intensity, mostly occurring in the shrublands. On the other hand, high intensity fires (22%) were generally distributed within Zimbabwe’s eastern and western regions. The third objective focused on detecting long-term spatiotemporal fire patterns in Zimbabwe using MODIS fire location data and a spatially explicit method (Emerging Hot Spot Analysis). The study also statistically analysed how the spatiotemporal distribution of vegetation fires is related to environmental factors. The research findings show that the occurrence of vegetation fire varies with seasons with the highest number of fires occurring in September. New information unveiled from the third objective indicated that fire activity tends to be high in June, July, and November despite these months being excluded from the official fire season in Zimbabwe, generally observed from August to October. Persistent, diminishing, oscillating, and historical spatiotemporal fire hotspots were observed in the northern regions of Zimbabwe. The final objective assessed the various topographic, bioclimatic, topographic, vegetation and anthropogenic factors that influence the occurrence of fires in Zimbabwe. The fire hazard levels were also predicted using the Maxent model based on the analysis of MODIS fire location data combined with topographic, bioclimatic, topographic, vegetation and anthropogenic factors. The jack-knife test evaluated the contribution of each variable towards the performance of the model, while the AUC (receiver operating characteristic curve) was used to estimate the model's accuracy. The research findings identified temperature annual range, precipitation seasonality, human influence and elevation as contributing highly to the occurrence of vegetation fires across Zimbabwe’s landscapes. The average AUC of 0.77 demonstrated good model accuracy. Conclusively, results from this thesis reveal the utility of spatial statistics and machine learning methods based on satellite fire data to understand spatiotemporal patterns of fire in Zimbabwe. Specifically, the detection of spatiotemporal patterns of vegetation fires, fire intensity clusters and their predicted hazard levels were successfully mapped. The information derived from this study is valuable in improving fire management in Zimbabwe and other regions. The detection of spatiotemporal patterns of fire, fire intensity and fire hazard levels result in new valuable information important for the mplementation of key fire management policies and strategies in Savanna ecosystems.