The development of a temporal-BRDF model-based approach to change detection : an application to the identification and delineation of fire affected areas
Although large quantities of southern Africa burn every year, minimal information is available relating to the fire regimes of this area. This study develops a new, generic approach to change detection, applicable to the identification of land cover change from high temporal and moderate spatial resolution satellite data. Traditional change detection techniques have several key limitations which are identified and addressed in this work. In particular these approaches fail to account for directional effects in the remote sensing signal introduced by variations in the solar and sensing geometry, and are sensitive to underlying phenological changes in the surface as well as noise in the data due to cloud or atmospheric contamination. This research develops a bi-directional, model-based change detection algorithm. An empirical temporal component is incorporated into a semi-empirical linear BRDF model. This may be fitted to a long time series of reflectance with less sensitivity to the presence of underlying phenological change. Outliers are identified based on an estimation of noise in the data and the calculation of uncertainty in the model parameters and are removed from the sequence. A "step function kernel" is incorporated into the formulation in order to detect explicitly sudden step-like changes in the surface reflectance induced by burning. The change detection model is applied to the problem of locating and mapping fire affected areas from daily moderate spatial resolution satellite data, and an indicator of burn severity is introduced. Monthly burned area datasets for a 2400km by 1200km area of southern Africa detailing the day and severity of burning are created for a five year period (2000-2004). These data are analysed and the fire regimes of southern African ecosystems during this time are characterised. The results highlight the extent of the burning which is taking place within southern Africa, with between 27-32% of the study area burning during each of the five years of observation. Higher fire frequencies are exhibited by savanna and grassland ecosystems, while more dense vegetation types such as shrublands and deciduous broadleaf forests burn less frequently. In addition the areas which burn more frequently do so with a greater severity, with a positive relationship identified between the frequency and the severity of burning.