In Part 1, we used satellite imagery and ground surveys to assess the impact of the relatively small, low intensity early season wildfire SUD008 in Central Ontario. Subsequently, we captured RGB imagery of the study area on May 30, 2025 using a UAV (DJI Mavic Mini 2, SZ DJI Technology, China) to an approximately 1.5 cm pixel size, which was later resampled to 20 cm to match the available RGBNIR aerial imagery from the Central Ontario Orthophotography Project (COOP) captured in Spring 2021. As a check of our previous use of difference Normalized Burn Ratio (dNBR), we visually compared whether the areas suspected of burn (surpassing dNBR thresholds of 0.1 and 0.2), lined up with signs of burn (Figure 1), where generally this was the case.

Figure 1:  Sentinel-2 based dNBR calculated from pre- and post-fire imagery, displayed as pixels exceeding typical thresholds of 0.1 and 0.2 in orange and red respectively, overlaid over high resolution UAV imagery captured on May 30, 2025.

Next, as a first pass to gauge our ability to scale up burn observations, we performed an unsupervised iso-cluster classification on the UAV imagery. This tool is available in the Classification Wizard in ArcGIS Pro (ESRI, USA) as quick means of classifying pixels of a raster into a user-defined number of classes. In this instance, we classified the UAV imagery into five classes, which were interpreted as two burned classes, two vegetation classes and open rock (Figure 2). The accuracy of this classification, or rather the ability to rapidly discern between burned and unburned was assessed using the DOB survey points as ground-truthing data, where the two burned classes were considered in agreement if ground-truthed as burned or singed and the three other classes with unburned.

Figure 2: Unsupervised iso-cluster classification of RGB UAV imagery captured post-fire on May 30, 2025, with inferred land cover classes.

The rapid, unsupervised classification yielded mixed results in discerning burned area, with moderate (0.62) producer accuracy and high (0.88) user accuracy, compared to moderate (0.76) producer accuracy and low (0.41) user accuracy for unburned. This may be attributable to the patchy distribution of singed ground and the lack of spectral difference between bare rock and burned ground, given the limitation of the UAV to only the visible colour bands.

As another method, the DOB survey points were used to train a supervised, random forest model to classify the UAV imagery based on burned status. After resampling the imagery to 20 cm for processing speed and the GPS accuracy of the survey points, a random forest model was trained to the UAV imagery and achieved moderate accuracy (0.75) and sensitivity (0.66 and 0.78, respectively) in discerning burned and singed pixels from unburned ones, based on a 20% sample of the survey points held back for validation. This was still fairly rapid, in a runtime of about a minute. Initial tests yielded poor accuracies when trying to distinguish between all three burned states separately so this model was only to simply distinguish burned and unburned pixels.

Within a study area covered by the UAV imagery captured and the DOB survey (7.86 ha), we then compared the burned area estimated by all three means: Sentinel-2 dNBR, unsupervised and supervised random forest classification based on the ground surveys and UAV imagery. Based on Sentinel-2 dNBR at thresholds of 0.2 and 0.1, there was an estimated burned area of 0.80 to 2.68 ha. The unsupervised classification of the high resolution UAV imagery provided a slightly higher estimate, based on the two classes interpreted as burned, of 2.97 ha. The random forest classification, which was inclusive of areas both burned and singed on the ground, gave the most generous estimate of burned area at 5.13 ha.

While our micro-drone was able to quickly capture post-fire imagery at high spatial resolution, the lower spectral resolution of just the three visible bands limited our accuracy in using image classification to scale up our burn survey measurements, though this was still at least moderate in most cases. This resulted in considerable variance in our estimates of area burned across methods. Next, we will investigate the potential factors, such as topography that impact the ignition risk.