Published works

Wildfire surveillance is an important aspect of e↵ective wildfire management, requiring near continuous observations to detect and monitor fires. Geostationary satellites have the potential to meet this challenge, capturing full disk images every 10 to 30 minutes at ground sample distances down to 500 m for some sensors. However, the MIR (Middle Infrared) and TIR (Thermal Infrared) channels on geostationary satellite sensors have a coarse ground sample distance of 2-4 km. Currently, fire detection algorithms depend on these channels to detect thermal anomalies. The coarse spatial resolution in the MIR and TIR channels limits the application of geostationary satellite for wildfire surveillance. This thesis looks to fully exploit the potential of geostationary satellites for wildfire surveillance through a multi-spatial and multi-temporal approach.

The first research question in this thesis, develops and tests an algorithm to improve the wildfire surveillance capabilities of the geostationary satellites. The new algorithm utilises the MIR, NIR and visible channels, linking them to biophysical processes on the ground. The MIR channel is used to detect thermal anomalies, the NIR channel is used to detect changes in vegetation cover, and the visible channel detects smoke from the fire. By combining these detections, or observations, fire surveillance can be achieved at the highest ground sampling resolution available (typically in the visible wavelength channels). Initial algorithm development and testing were conducted on the Advanced Himawari Imager (AHI) sensor onboard the Himawari-8 satellite. The MIR, NIR and RED channels on AHI have 2 km, 1 km and 500 m ground sampling distances respectively, enabling the new algorithm to detect 2 km thermal anomalies and 500 m fire-line pixels. Fire-line pixels is a new product designed to detect the trailing edge of the fire.

Quantifiable methods for assessing algorithm performance in geostationary satellites are dicult to apply due to their high temporal resolution and lack of concurrent in-situ information. The second research question investigates methods for assessing the performance by considering the near continuous temporal sampling of geostationary satellites and the higher spatial ground sampling resolution a↵orded from LEO (Low Earth Orbiting) satellite observations. The study examines di↵erent evaluation methods and suggests a three-step process to provide the optimum performance evaluation for geostationary wildfire surveillance products, inter-compared with LEO satellite-based thermal anomaly detections.

Algorithm performance is further evaluated in research question three using the intercomparison method developed in research question 2 and applied to case study fires over Northern Australia. Subsequently, the algorithm is evaluated using an annual dataset (2016) comprising of nine study areas across Australia (totalling 360.000 km2) stratified by tree canopy cover. The algorithm reported an omission error of 27 % at 2 km ground resolution when compared to VIIRS (Visible Infrared Imaging Radiometer Suite) hotspots over the nine study grids. In Northern Australia, the algorithm detected fires up to three hours before LEO observations due to the high temporal frequency of observations. Furthermore, in comparison to MODIS (Moderate Resolution Imaging Spectroradiometer) hotspots, there was a 73 % chance of detecting fire activity at the location of the MODIS hotspot, before the MODIS overpass. The algorithm also demonstrated a 40 % detection probability for fires less than 14 ha over Northern Australian woodlands. The fire-line pixels with a ground sampling distance of 500 m demonstrated a 25 % commission error when compared to VIIRS hotspots over the nine study grids. Over Northern Australia, this figure was 7 % inter-compared to Landsat-8 burnt scars.

The fourth research question applied the developed algorithm to the SEVIRI (Spinning Enhanced Visible and Infrared Image) sensor onboard the European Meteosat Second Generation (MSG) satellite. SEVIRI has an operational fire product (FIR (Active Fire Monitoring)) which provides 3 km ground resolution hotspots using the MIR and TIR channels. The algorithm initially developed for AHI was modified to work with SEVIRI 3 km MIR channel and the High-Resolution Visible (HRV) channel (1 km). An inter-comparison of the modified algorithm with FIR products showed a 28 % and 16 % improvement in commission and omission errors respectively over a large case study fire in Portugal. The modified algorithm also improved the SEVIRI wildfire surveillance ground sampling resolution to 1 km taking advantage of the HRV channel.

The algorithm developed in this study demonstrates a novel approach to utilise geostationary satellites for wildfire surveillance with improved spatial resolution. Compared to the 2 km thermal anomaly hotspots derived through existing algorithms for AHI, the new algorithm provides 2 km thermal anomaly detections and 500 m fire-line pixels with performance comparable to that of medium resolution LEO satellites. Near-real time implementation of the algorithm has the potential to provide high temporal fire surveillance capabilities. The fire-line pixels from the algorithm could also be used to derive fire behaviour parameters such as heading and speed, providing an essential tool for wildfire surveillance in remote parts of Australia and other areas, where resources can only be deployed for a hand full of high-risk fires.