Entwicklung einer Prozesskette zur automatischen Detektion von Brandflächen auf der Basis von Sentinel-2 Daten

Kurzfassung

Wildfires not only threaten human lives and livelihoods, but also lead to enormous economic and environmental damage. Annually, biomass fires burn a total area of 3.5 - 4.5 million km² worldwide, leading to large amounts of greenhouse gases being released into the atmosphere. In consequence the fires have a significant impact on global warming. Especially due to the positive feedback created from further dehydration of the forests and in turn, a further increase in fire risks. In order to estimate the extent of the damage and to model the impact on the climate, as well as for planning in forestry and fire management, it is essential to have a precise knowledge regarding the location, the extent and the frequency of the fire incidents. Remote sensing provides a unique opportunity for this objective, as the earth's surface can be observed large-scale and with high frequency. Because of their high spectral, temporal and geometric resolution, the data from the new Sentinel-2 mission has the potential to enable this task. In particular, the simultaneous observation using two identical satellites makes it possible - despite the geometric resolution of 10 m - to achieve repetition rates of 5 days and less. Therefore this thesis aims to develop a process for the automatic detection of burnt areas based on Sentinel-2 data. First of all it is necessary to prepare the data for the fire detection. The preprocessing includes a coregistration, atmospheric correction and also a format conversion. In the subsequent step a method for cloud detection is developed. This serves to identify regions to the user, which carry no information on the existence of burnt areas. It also allows a reduction of computation time, as these areas are eliminated in the further processing. Finally the detection of the burnt area can be carried out. Based on an examination of the procedures and criteria from other methods, a new approach is developed. A discriminant analysis is performed to examine mono- and multitemporal features. This enables the determination of features, which are suitable for the identification of the burnt areas. For the detection of the burnt scars, a two-phase algorithm is developed. The first phase is used for a threshold based generation of seed pixels in the burnt areas. To avoid false detections, conservative criteria are empirically determined based on training data. In the second phase the neighborhood of the seed pixels is examined with more liberal criteria. For this purpose, a region growing method is combined with a Support Vector Machine (SVM) classifier, wherein the seed pixels from the first phase are used as training data. In order to estimate the quality of the process, a validation is carried out. This shows that the method has a high sensitivity for detecting the burnt areas. However, it is also identified that the algorithm is not able to separate burnt areas reliably from harvested agricultural land. Furthermore, its capacity to recognize burnt areas in cloud shadows or to detect dark burnt areas is limited.