The use of radar has been proven for forest fires, but moorland fires result in a less significant change in habitat structure. Notably, in England, it is recognised that it could be hard to distinguish between heather burning and heather cutting, but the critical thing is to identify
the change; it may be that optical data would be used to confirm the type of change.
In response to a Department for Environment & Rural Affairs (DEFRA) Invitation To Tender, a six-month research project focused on developing techniques for identifying burn scar areas from the Sentinel-1A and -1B satellite data. Workflows were implemented in Jupyter Notebooks (available from https://github.com/pixalytics-ltd/upland-burn-detection) shared within the team via a Jupyter Lab instance. The notebooks could be run individually or in sequence to downloaded, pre-processed and applied detection algorithms; the coherence processing was undertaken using ESA's SNAP Toolbox called via its GPT interface. The outputs from the algorithms were then compared to imagery for known burn areas, and discrepancies were investigated so that an iterative approach could be used to develop the final algorithm. Three case study areas were used for testing and analysis: Isle of Skye and Eastern Cairngorms in Scotland and England's Peak District National Park (PDNP).
Results showed that burn scars were visible in coherence data. Burn areas showed low coherence in image pairings that covered the burn date, followed by high coherence in the images that followed, presumably due to lack of vegetation/growth. In fact, in some cases, the burn scar was still visible the following year.
The next step was to develop an automated detection algorithm that could be used to process time-series datasets. However, coherence data are normalised between 1 and 0 for each image, which introduces a potential hurdle as images will all have the same scale range regardless of their relative coherence. Therefore, the first step was to reverse this normalisation. Although this is impossible to do correctly without the initial maximum and minimum values used to normalise the image, the effect could be replicated by dividing each image by its median value. Instead of looking at the difference between consecutive images, an average coherence value was created for every pixel across the date range used. Then, the absolute difference between each image and the temporal mean image was calculated before all images were summed. A threshold was applied to extract the most changed pixels likely to be burn areas.
In summary, the findings suggest that burn areas' detectability with coherence improves over one year following fire. However, further ground-truth work on post-fire regrowth is needed to understand the mechanisms responsible for this response. The Jupyter Notebooks have been made publicly available and will continue to be developed for this application alongside being reused in future collaborative projects.