Abstract
Wildfires play an important role in regulating and maintaining grassland ecosystems
by, for example, supressing bush encroachment and the impacts of over grazing.
However, they still present risks to the natural welfare of human beings, especially if
they run out of control. In the short term, wildfires can destroy vegetation communities
and disrupt ecosystem services. Greenhouse gases such as carbon dioxide are
released by the wildfires; these gases can have long term effects on human health
and lead to climate change. Despite the grassland biome being one of the largest
biomes influenced by wildfires, it does not receive as much attention. The increasing
frequency of wildfires prompts the need for proper monitoring techniques that could
help to reduce negative impacts from fire within the grassland biome. This study aimed
to propose techniques based on remote sensing to monitor and predict fires in the
grassland biome in the Mangaung Municipality of the Free State (a province in South
Africa). To achieve this aim, three objectives were set. Objective 1 sought to
investigate the efficiency of optical and radar remote sensing in mapping burn scars
in the grassland biome. Objective 2 sought to assess the spatiotemporal burnt area
dynamics, namely the spatiotemporal patterns, using remote sensing data. Objective
3 set out to assess the utility of vegetation characteristics derived from remotely
sensed data in the prediction extent of burnt area. To address Objective 1, the Random
Forest (RF) and Support Vector Machine (SVM) algorithms were used to classify
Sentinel-2 optical and Sentinel-1 radar data in Google Earth Engine. The optical data
produced accuracies of over 90%, showing its better performance in distinguishing
burn scars from other classes (namely water bodies and grass). In contrast, radar data
produced low accuracies (less than 50%), which can be explained by the confusion
between classes. When the radar and optical data were combined, the accuracies
were similar to that of the optical data. Notably owing to the radar data’s ability to
distinguish burn scars from shadows, discrimination of burns scars from shadows was
better in this instance. To achieve Objective 2 a space-time cube was created by using
32-years of Burn Area Index (BAI) data derived from Landsat imagery. The burn
patterns were analysed using the emerging hot spot analysis which showed that there
were oscillating cold spots and oscillating hot spots. The oscillating hot spots showed
that there were historical trends, however recent years changed to be cold spots. To
ii
further explore these patterns, two predictive models were applied to the cube, namely
the curve fit and the RF, with the results showing a better performance for the RF. To
achieve Objective 3, three different datasets in 2007, 2014, 2021 were used with two
images from each year. The first image was taken around a month before the burn
and the second after the burn. The Multiscale Geographically Weighted Regression
(MGWR) was used on optical vegetation indices extracted from the first image taken
prior to the burn. The indices encompassed the Enhanced Vegetation Index (EVI), the
Normalised Difference Vegetation Index (NDVI), the Moisture Stress Index (MSI), and
the Normalised Difference Moisture Index (NDMI). The results showed that while there
was an overestimation in the size of areas that would burn in the future, there was a
correlation showing that the extent of burnt area could be predicted using vegetation
indices. This study contributes to the creation of management and mitigation strategies
that can reduce fire damage and prevent the future destruction of the environment in
the grassland biome.