methods for background temperature estimation in the context of active fire detection
Bryan Hally is a PhD student in the department of Natural Resources (NRS). His supervisors are prof.dr. A.K. Skidmore from the faculty of Geo-Information Science and Earth Observation (ITC) and prof.dr. S. Jones from RMIT University.
Fire is an integral catalyst for change and regeneration in the environment, along with being a major impact on social and economic activity. In evolving climatic conditions, wildfires are increasing both in intensity and in area impacted in recent years. Remote sensing has been used for many decades to provide insight into fire activity and impact through the use of infrared imagery for active fire detection. Electromagnetic wavelengths at around 4 μm are particularly sensitive to fire activity, in comparison to nominal conditions which incorporate both solar reflection and thermal emission of the earth’s surface. Brightness temperature measurements in these wavelengths isolating such radiative anomalies enable fire detection and attribution from satellite sensors. The processes involved in providing these active fire products require an accurate estimation of the background brightness temperature of the area in question without the influence of fire. These commonly used active fire products generally use a contextual-based estimate to provide this background temperature information. Whilst this estimation technique is widely accepted for use, especially in single time-point polar orbiting products, the introduction of new generation geostationary sensors provides substantial improvements to knowledge about the earth’s surface, especially with regard to diurnal variation. These sensors provide an opportunity to not only evaluate the accuracy of context-based estimation of brightness temperature, but to integrate the rich spatio-temporal information provided by such sensors to improve the accuracy and availability of background estimates.
In order to determine an adequate level of accuracy required for the derivation of new temperature estimators, it was important to know the accuracy of the current paradigm of contextual temperature estimation. To date, whilst contextual temperature estimation is widely used, no definitive study of the expected error in temperature estimates had been completed. An analysis of the error involved in contextual estimation was conducted upon medium wave infrared radiation (MWIR) images taken by the AHI-8 sensor onboard the Himawari-8 geostationary satellite. Comparisons were made between contextual estimates and the raw brightness temperature observations over the AHI-8 full disk for 36 images at 0500 UTC across 2016, and across a number of case study areas for 31 days of images surrounding the peak fire period as determined by the VIIRS active fire product. The study found that variation in temperature estimations from context had negligible bias and standard deviations around 1.1 K when the surrounding 5_5 area was clear of cloud, which occurred in 53.9 % of cases. Accuracy diminished as the contextual estimation surface was obscured, such that pixels with 65 % or more context available experienced a 56 % increase in estimate variation. The common practice of window expansion saw the variation of estimates increase substantially, with 7_7 windows resulting in a 44 % increase in variation over the 5_5 results. The study concluded that 5_5 contextual estimation should be limited to using values where at least 65 % of the contextual surface is available, with no expansion of the contextual window due to the detrimental effect on estimation accuracy. This resulted in 1 in 7 non-obscured pixels (14.5 %) in the examined images not having accurate contextual estimates available. Common causes of increased contextual inaccuracy included coastlines, land cover changes, slope and aspect of the surface, urban heat signatures and land inundation, with these effects highlighted in the examined case studies.
With the identification of inadequacies in the contextual estimation method, investigation of methods to leverage the temporal domain of the geostationary sensor to fill these gaps was undertaken. Particular focus was placed upon the modelling of the diurnal temperature variation of locations, and in particular how gaps in the training data for such models could be filled. Previous studies that had used diurnal modelling in typically cloud obscured areas had identified deficiencies in the use of single pixel data for creating temperature models. A new technique was developed using a standardised model of diurnal temperature variation based upon the latitude of the examined area, and corrected for local solar time. Results from models created by this method, known the Broad Area Training (BAT) method, were compared to a single-pixel derived model and the raw temperature recordings from AHI-8. The comparison found that the RMS error of the BAT-derived models maintained sufficient accuracy for temperature estimation with up to half the estimation days’ values obscured by cloud, with errors reduced by more than 50 % compared to the single pixel method with between 30 – 70 cloud affected images present in the day of estimation. The method also increased the availability of training data for modelling using this type of multi-temporal method, with up to 90 % of pixels across the Australian continent possessing sufficient training data for estimation, in comparison to 40 % for the single-pixel model.
The success of brightness temperature estimation using the BAT method led to an investigation of the potential isolation of brightness temperature anomalies using these models. A comparison study was undertaken using active fire information taken from the MODIS and VIIRS active fire products, and burned area information from the TERN Auscover MODIS Burned Area product, and information from these three sources was compared to anomalies isolated from the BAT modelling of the affected locations using a number of rudimentary temperature thresholds. Anomalies were detected the BAT modelling of temperature against raw image temperatures in between 75–99 % of cases where a LEO fire detection took place, with variation based upon the threshold set. Synchronous fire activity was detected between LEO fire products and the BAT anomalies in between 46–68 % of case where fires were identified by both active fire products. Using BAT to find anomalies also resulted in an increase of anomalies detected above the LEO products, with between 50–75 % of burned area pixels without LEO active fire hotspots resulting in anomalous temperature detections. The comparison of anomaly time of detection versus the LEO fire products was not unexpectedly favourable to the BAT anomalies, with improvements in initial isolation of 5–7 h over the two LEO products used.
Whilst the BAT method of temperature modelling was relatively successful in isolation of anomalies in geostationary imagery, investigations led to the identification of methods of estimation that could theoretically be applied to both geostationary and LEO sensor imagery. A new method of background temperature estimation was developed, this time using the similarities of temperatures measured in a search radius around the target to be estimated, in order to derive suitable candidate pixels for estimation from a single image. This method, known as the Spatio-Temporal Selection (STS) method, was applied to images from a number of case study areas across the AHI-8 full disk, and comparisons were made with values from the prediction image and the 5_5 contextual estimation. The STS method demonstrates a 10–40 % improvement in variation over contextually derived temperatures, with marked improvements in visually assessed accuracy. The method also provides more estimates of temperature than the contextual estimator, with between 16–45 % more pixels able to have their brightness temperature estimated. The study demonstrated the potential extension of the method into use of LEO imagery and highlighted other deficiencies with the contextual estimation method that the first study of the thesis did not identify.