Sentinel-2 Fire Scars - DETSI Algorithm, QLD Coverage 

Summary: Fire scars were automatically detected in Sentinel-2 imagery using differenced bare soil fraction values obtained using the Joint Remote Sensing Research Program’s (JRSRP) fractional cover model. A Sentinel-2 pixel is identified as likely burnt if there has been a significant increment in bare soil fraction relative to the previous fractional cover values. Once areas of likely change are detected, a region growing algorithm is applied to expand the area to capture the whole fire event. Areas are then filtered into burnt and unburnt classes using a decision tree analysis. Subsequent manual interpretation is used to delineate between false positives and true positives. 

RapidFire algorithm: The RapidFire algorithm is applied to the JRSRP Sentinel-2 fractional cover product. The RapidFire algorithm identifies core pixels of potentially burned area, based on the temporal difference in bare soil cover fraction. Core pixels are spatial clusters (bigger than 15 pixels) where the change in bare cover fraction exceeds an optimised threshold. The extent of the pixels classified as potentially burned are expanded, using a region growing algorithm on the core pixels. An object-oriented classification is used to discriminate between burned and unburned areas. The classification tree was based on the median values of the temporal difference of NBR (dNBR) and NIR + IR (dNIRIR) of each potentially burned area. 

Manual Editing: Manual editing is conducted by trained analysts to reduce the number of false fires and omission errors. 

Monthly Composite: Single date manually-edited fire scar data is composited into monthly periods. 

Annual Composite: All mapped fire scars across QLD in a January to December period are composited together to create annual composites. 1-12: month (of Sentinel-2 acquisition) when fire scar was first detected; 254: crop/water masked (using Current Queensland Land Use Mapping) - no fire scar detection conducted. 

Consistency: Sentinel-2 analysis does not provide a complete record of fire history. Fire scars may be missed or under-mapped due to: 1) Lack of visibility due to cloud, haze and smoke, and cloud shadow; 2) Misclassification as non-fire related change or cloud shadow; 3) Lack of detection due to size or patchiness. Fire scars smaller than 2 ha may not be included; 4) Lack of detection due to rapid regrowth of vegetation. This is particularly an issue when there have been multiple cloud-affected images in the time series; 5) Lack of detection for cool grass/understorey fires, obscured by unburnt vegetation; False burned areas or over-mapping may result from: 1) Omission errors in the cloud/shadow masks, where cloud is classified as fire scar; 2) Areas of high intensity land-use change where the extent of bare ground increases rapidly (e.g cropping, vegetation clearing); 3) Areas of inundation (e.g tidal flats, wetlands, ephemeral lakes and channels). 

Attribute Accuracy: Fire scars may persist and continue to be detected for several months in the image time sequence. Where there has been fire scar persistence or multiple fire scars recorded for a given pixel within the compositing year, the earliest month of detection is recorded. 

RapidFire Approach: This approach has some important consequences: 1) Not all the pixels of an image are analysed due to cloud and shadow effects; 2) Time elapsed between observations for different pixels of the same image may differ, again due to cloud and shadow effects over time; and 3) Burned areas only appear once in the record. If for some reason a burned area is missed in the first unmasked observation it will be missed in the whole record. This new method is a different approach from the previous Landsat-derived fire scar mapping program (1987-2016). That automated time series method identified large negative outliers in reflectance indices (based on NIR and SWIR1 bands) relative to the time series.