Introduction

In the recent past, wildfires of unprecedented scale and intensity around the world have made their way into the minds of most, often accompanied by concerns regarding climate change. With anthropogenic activities driving warmer temperatures around the globe, the likelihood, severity, frequency, and longevity of wildfires is expected to rise. According to the United Nations Environment Programme, "[e]ven with urgent action, the number of wildfires globally is expected to increase 50 per cent by the end of the century" (1). In our modern technological era, comprehensive, accurate, and reliable methodologies for conducting geospatial analyses on regions impacted by wildfires are increasingly being employed to inform and guide communities, governments, and international organizations in wildfire mitigation, firefighting, as well as rescue and evacuation efforts. 

This past Summer, I helped to develop a new version of the Bayesian Updating of Land Cover (BULC) algorithm (2) – an algorithm capable of accurately visualizing wildfire burn-scars from satellite imagery – originally developed by Dr. Jeffrey Cardille of McGill University's Department of Natural Resource Sciences (3). The algorithm uses the Normalized Burn Ratio (NBR) spatial index – an index that capitalizes on the normalized difference between NIR and SWIR bands – to identify "burned" pixels in an image, where a pixel is considered to be burned when its NBR value – ranging from 0 (darker) to 1 (brighter) – drops below a certain threshold (4). BULC then compares NBR values in pixels from images taken in a year that the region experienced a wildfire to pixels in images taken in a recent year previous that did not. The difference is taken between the target year (i.e., saw fire) and the expectation year (i.e., did not see fire) NBR values and drastic drops in the same pixel over multiple temporally successive images are flagged as burned. The most recent version of the algorithm  – BULC Version D (i.e., BULC-D) – is capable of accurately detecting changes in NBR while accounting for natural variations in the index due to yearly changes to the land cover, such as seasonality. By employing a harmonic regression to adjust the mean index value depending on the time of year, Version D of the algorithm can account for "false burns", pixels where natural phenomena – like changing foliage color in deciduous forests in autumn, changes in vegetation cover due to drought, and deforestation – have historically been incorrectly flagged as burned.

Project Objectives and Questions

In theory, it is possible to re-task the BULC-D algorithm to use another spatial index, given that some modifications to the underlying script are made. As such, I propose a project aimed at adapting the BULC-D algorithm to work with the Burned Area Index (BAI) – an index that detects burn scars using the spectral signature of charcoal in post-fire images (5).

This project is aimed at repurposing the BULC-D algorithm to utilize the Burned Area Index (BAI). The following objectives will be pursued with the ultimate goal of delivering a new "spectral lens" from which wildfire burn scars can be analyzed, hopefully increasing the overall accuracy, flexibility, and usability of the algorithm:

Objectives

(1) Find 4 test fires that burned sometime in the past 10 years – with corresponding official fire progression maps – that was not accurately visualized using BULC-D informed by NBR.

(2) Rewrite the NBR-informed BULC-D script in GEE to instead rely on BAI for its computations. This will likely involve a variety of changes to the algorithm on multiple levels.

(3) Compare BULC-D's NBR output to its BAI output for each of the 4 test fires. Both outputs will then be cross-compared with the official fire progression maps to see if one seems to perform better than the other overall.

Questions

• How well does a BAI-informed BULC-D visualize burn scars as opposed to official fire progression or extent maps of the same fire?

• How does BAI-informed BULC-D perform compared to its NBR-informed counterpart in:

  • • regions where NBR-based burn scar visualization has performed well?

    • regions where NBR-based burn scar visualization has not performed well?

Methodology

Objective 1: Identifying test fires

Four candidate test fires (detailed in Data section, Candidate Test Fires) were identified based on project objectives and questions. These fires were chosen because: 

• all fires occurred within the past 10 years

• all fire accounts are accompanied by one or more official fire progression or extent maps

NOTE: a map is considered to be official if it has been developed by a known government or organizational entity, ensuring at the very least that the contents of the outsourced map have a higher chance of having been cross-validated and peer-reviewed before dissemination to the public. This necessary step ensures that comparisons, as well as any determinations based on those comparisons, are being performed using reliably outsourced data.

The first step was to identify fires that have met the criteria specified in the above section (i.e., Data, Candidate Test Fires). Test fires outlined in this proposal are candidates for the project and may be swapped if one or more appropriate alternatives are found. These fires were found by searching the web for information on the location, start time, end time, season (i.e., Spring, Summer, Fall, Winter), and burn extent. Other information regarding environmental conditions (i.e., land cover type, topographical similarities, meteorological similarities, seasonal similarities, etc.) were kept in mind as important qualifying characteristics for candidate fires. The 2020 Cameron Peak and East Troublesome fires in Colorado – fires that were originally identified as having an NBR-informed output that disagreed with the accompanying official fire extent maps – were chosen as two of the test fires for this project (i.e., candidate fires 1 and 2). Next, a 2020 fire that took place in Ukraine's Chernobyl Exclusion Zone – a region where environmental conditions were relatively similar to those of the Colorado fires – was chosen (i.e., candidate fire 3) in an attempt to identify common conditions that could be affecting the NBR-informed output. Lastly, the 2019-2020 Kangaroo Island bush fire– a fire with conditions similar to regions where the NBR-informed output has historically performed relatively well (i.e., arid/semi-arid regions, such as Southern California) – was chosen (i.e., candidate fire 4). 

Each of these fires were validated by cross-comparing the NBR-informed BULC-D output with the fire's official progression or extent map. Evaluations on whether the BULC-D output agreed or disagreed with the outsourced map were performed manually through visual confirmation. It is important to note that this project is not currently relying on formal statistical analyses to guide our understanding of whether an NBR-informed output and a corresponding map agree or disagree. At least at this moment, judgements are being made based on observation alone as it seems to not only be too time consuming and beyond the scope and scale of this project to incorporate statistical determinations in the results, but validation via visual confirmation has thus far yielded satisfactory results since the inception of my Honours Project. 

• • Potential Deliverable: visualization(s) comparing outsourced official fire progression maps to the NBR-informed BULC-D output. The NBR-informed BULC-D output of these fires will be verified as being relatively inaccurate, with most outputs being too "noisy" or dissimilar from the official maps for our liking.

Objective 2: Modify BULC-D to take BAI as its input index

Once we have chosen our four test fires, the next step is to modify the BULC-D algorithm to take Burned Area Index (BAI) as its input index instead of the Normalized Burn Ratio (NBR) index. Re-tasking the algorithm to use BAI as its input index is not as straight forward as it may seem for several reasons. First, we have to understand how BULC-D works, which requires we understand how the algorithm takes in and uses satellite imagery to calculate an NBR value for each pixel in each image of a given image collection. BULC-D takes in image collections from both Landsat 8 and Sentinel 2 satellites; these chronologically organized year-long catalogues of satellite imagery are utilized by a variety of functions within the BULC-D framework to calculate a given index, such as NBR or BAI. Each index is calculated by performing a specific set of ordered operations on the spectral bands required to actually calculate the index. Each spectral band represents a different range of wavelengths of energy at a variety of spatial resolutions and can be used to selectively filter an image for natural phenomena happening on the surface.  In the case of NBR – which uses the normalized difference of Near Infrared (NIR) and Shortwave Infrared (SWIR) bands (equation below) – the spectral signature of vegetation is reflected strongly in NIR and scarred/blackened woody vegetation and earth are reflected strongly in SWIR.

Using this equation, an NBR value is calculated within each pixel of each image in a given image collection according to the schematic below. First, an image is singled out from the Landsat 8/Sentinel 2 Image Collection (IC) and the NIR and SWIR bands are isolated based on their band names. Band names for NIR and SWIR in L8 and S2 differ (11), so conditional "if" statements (on the right below) are used to select the bands by their respective names. Next, an in-script expression is written, representing the normalized difference of these bands. At this point, the index has been calculated and the resulting NBR value is added to the image as a new band called "NBR" and the image is added remerged into the IC until a full year's worth of image is aggregated.

As has been mentioned previously, once a given index has been calculated and attached to an image, it is computed for every pixel in that image; this is the case for every image in the IC. Keeping in mind that each IC represents a year's worth of images, this presents a unique opportunity to make intra-IC comparisons, effectively allowing for the observation on changes in a given index over the entire year (i.e., changes between images). For burn scar detection, however, more interesting is the potential to make inter-IC comparisons – meaning the ability to make observations on changes in a given index between one year and another (i.e., changes between ICs). This is the central tenet behind how BULC-D is able to detect burn scars. By taking the difference between index values from a known a posteriori fire image at year i+1 and values from a known a priori non-fire image at year i, pixels with significant enough inter-annual variability to suggest a change in the underlying land cover (e.g., a fire) since the non-fire year are flagged as changed.

These changes are standardized using a Z-Score (equation below), which subtracts the sample value (x) from the sample mean (µ) and divides the result by the standard deviation (σ).

On its own, each collection bin will not accurately capture the annual variability exhibited by a pixel such as this.  So then what can be done when multiple images of the same pixel have been placed in multiple bins? Alternatively, what would happen if the last image of your IC is in bin 10 but the rest are in bin 5 and 6?  There are cases when the final state of a pixel at the end of the observation period is not always representative of the actual conditions on the ground. Ideally, when attempting to identify a real change on the surface, there should be continuous, consecutive evidence of change over multiple images. For instance, BULC-D should not be absolutely certain that a pixel is burned if it has only dropped into bin 2 once throughout the year. Rather, we need to continually update BULC-D's final decision class for a given pixel so that the final output is an accurate representation of the actual change that has taken place on the surface since the a priori non-burn year, without being swayed by passing non-fire phenomena like clouds or smoke. The algorithm accounts for this by feeding our information about the index value and the bin it has been placed in through Bayes' Theorem (equation below), which solves for the probability of a decision class (i.e., index either dropped, didn't change, or increased relative to the non-fire year) actually occurring given that it was placed in the correct collection bin. 

Now that we have covered how BULC-D works with NBR, we have to recreate the underlying code for BAI. As daunting a task as this sounds, most of the code can be replicated word for word and NBR replaced for BAI. Certain changes needed to be made, however, because of dissimilarities between the two indices. It was stated in the proposal for this project that changes would need to be made to the way in which the algorithm computed the Z-Score for BAI, but this is not the case. Instead, changes will be made to the algorithm's custom transition matrix. Because NBR and BAI span very different integer ranges (i.e., covered in Introduction & Background), fire and non-fire phenomena span different integer ranges and will be processed differently depending on the index one looks at. If the transition matrix is not adjusted according to the overall distribution of BAI, the algorithm will instead draw the BAI-informed output according to the distribution of the NBR index, yielding determinations about the burn state of each pixel based on the wrong index. 

As a solution, a separate transition matrix will be made for BAI to ensure that the likelihoods being fed to BULC-D are representative of BAI's distribution. Two randomly sampled datasets, each amounting to 30 data points (n = 30) of BAI values, will be created. Sample data will be tabulated in Excel and categorized as being either "burned" or "unburned" (i.e., n = 30 data points for each category). Burned and unburned pixels from one or more of the test fires will be plotted adjacently on a histogram and conclusions on their overall distributions will guide the development of a custom decision matrix for BAI. This will allow us to accurately capture the range of possible BAI values in a fire image and inform the algorithm on which BAI values to classify as burned.

After creating the new transition matrix and implementing the aforementioned code changes, the algorithm should now be correctly calculating and visualizing the BAI value within each pixel of each image in both the target year and expectation year image collections. At this point, we would be able to continue onto Step 3.

• Potential Deliverable: screenshot(s) and accompanying explanation of script, particularly parts of the code that were challenging/especially important to implement. 

Objective 3: NBR vs. BAI vs. Fire Progression Maps

Once we've sorted out the necessary changes to the code underlying the BULC-D algorithm outlined in Step 2, we can compare NBR-informed and BAI-informed BULC-D outputs to one another, as well as compare both to the official fire progression or extent maps for each of our four test fires. As in Step 1, each of these fires will be evaluated on whether the two BULC-D outputs agree or disagree with the outsourced map. These evaluations will be performed manually through visual confirmation. As was mentioned previously, evaluations via visual confirmation have yielded satisfactory results throughout my Honours Project thus far, leading me to believe that we can apply this method without any issues. 

• Potential Deliverable: visualization(s) comparing the new BAI-informed BULC-D output to its NBR-informed counterpart, as well as to outsourced fire progression maps of the 4 test fires.

Results

Objective 2: Modify BULC-D to take BAI as its input index

The process of computing BAI from Landsat 8 and Sentinel 2 imagery is extremely similar to NBR, except we instead use the BAI equation (below) and the appropriate spectral bands (i.e., Red and NIR) from L8 and S2 to compute the index value for each pixel of every image in the image collection.

Apart from these changes to the way in which BULC-D takes in spectral bands from Landsat 8 and Sentinel 2, all parts of the script pertaining to NBR being calculated and fed to BULC-D were duplicated for BAI with the only difference being the name of the index. In addition, 30 burned and 30 unburned pixels were randomly sampled from two test fires using the sample BAI data's raw values, mean, and standard deviation. These were manipulated with the NORMDIST(), ARRAYFORMULA(), FREQUENCY() functions available within Google Sheets and two histograms were created (seen below). 

Using these histograms as a baseline understanding of what the BAI distribution looks like in burned and unburned pixels, a custom transition matrix (bottom left) was created to manipulate BULC-D posterior probabilities with respect to this new index. The matrix is composed of numerical values represent relative likelihoods of each of the three final decision classes (i.e., BAI increase, no change, BAI drop). While testing these these binning likelihoods and their ability to capture true fire phenomena, these values were plotted (bottom right) to be able to visualize how the algorithm would assign a given pixel to one of the three final decision classes would and affected by each bin value returned by BULC-D. 

Objective 3: NBR vs. BAI vs. Fire Progression Maps

1) 2020 Cameron Peak Fire, Colorado

The NBR-informed (left) and BAI-informed (right) BULC-D outputs, as well as the official fire extent map (center below), for the Cameron Peak fire that took place in Colorado in Fall, 2020. The expectation (2019) and target (2020) year ICs range from fractional day 135 (May 15th) to 290 (October 17th). The day step size (i.e., the number of days between each image) is 2 days, meaning there is 1 image for every 2 days in the IC. Important supplementary information for the Cameron Peak fire that is pertinent to the results of this project includes the following:

• Incident time range: 8/13/2020 – 12/2/2020 (7)

• The Cameron Peak fire was the largest in Colorado history (6). 

• 8-14 inches of snow fell in the region on the evening of 9/8/2020, temporarily halting the fire (6).

• 8-18 inches of snow fell in the region between 10/24/2020 and 10/25/2020 (6). 

2) 2020 East Troublesome Fire, Colorado

The NBR-informed (left) and BAI-informed (right) BULC-D outputs, as well as the official fire extent map (center below), for the East Troublesome fire that took place in Colorado in Fall, 2020. The expectation (2019) and target (2020) year ICs range from fractional day 170 to 325. The day step size (i.e., the number of days between each image) is 2 days, meaning there is 1 image for every 2 days in the IC. Important supplementary information for the East Troublesome fire includes the following:

• Incident time range: 10/14/2020 – 11/30/2020 (7)

• E. Troublesome fire was the second largest in Colorado history (7).

• Snow fell in the region between 10/24/2020 and 10/25/2020, dramatically affecting fire behavior (7).

3) 2020 Chernobyl Exclusion Zone Wildfire, Ukraine

The NBR-informed (left) and BAI-informed (right) BULC-D outputs, as well as the official fire extent map (center below), for the Chernobyl Exclusion Zone (CEZ) fire that took place in Pripyat, Ukraine in Spring, 2020. The day step size (i.e., the number of days between each image) is 2 days, meaning there is one image for every two days in the IC. The expectation (2019) and target (2020) year ICs range from fractional day 1 to 156. Important supplementary information for the CEZ fire includes the following:

• Incident time range: 4/4/2020 – 4/14/2020 (8)

• The fire burned through the Red Forest, which is known to be radioactive (8)

• ~30% of tourist attractions were destroyed by the fire (8)

• Kyiv, Ukraine had the worst air pollution in the world one point on 4/16/2020 due to the fires (8)

4) 2019-2020 Kangaroo Island Bushfire, Australia

The NBR-informed (left) and BAI-informed (right) BULC-D outputs, as well as the official fire extent map (center below), for the Kangaroo Island bushfire that took place on Kangaroo Island, Australia in Winter, 2019 to 2020. The expectation (2019) and target (2020) year ICs range from fractional day 1 to 77. The day step size (i.e., the number of days between each image) is 1 day, meaning there is an image for every day in the IC. Important supplementary information for the Kangaroo Island bushfire includes the following:

• Incident time range: 12/20/2019 – 1/21/2020 (9)

• Fires burned approximately 49% of the island (9).

• The initial Duncan and Menzies fires were started by lightning strikes on 12/20/2019 (9).

• Also started by lightning strikes, the Ravine fire began on 12/30/2019 (9).

Discussion

Objective 2: Modify BULC-D to take BAI as its input index

Importantly, BAI and NBR display true fire phenomena in opposite ways, with NBR values dropping and BAI values increasing when a pixel is burned. BAI values also seem to peak as the fire incident is underway and the fire is actively burning and fall after several days. NBR, on the other hand, drops and stays dropped so long as vegetation growth or non-fire phenomena do not interfere. As a result, it seems that BAI reacts more sensitively to active burns, both peaking during incidence and falling after the fire has dissipated. This also suggests that any phenomena that stifles active burns, such as rain or snow, could disrupt BULC-D's ability to accurately determine a final decision class for BAI-informed outputs at true fire pixels.

Objective 3: NBR vs. BAI vs. Fire Progression Maps

Performance and ability to capture real fire activity varied between NBR and BAI at every one of the four test fires. For NBR, green pixels are representative of the "no change" final BULC-D decision class; red pixels are representative of the "NBR drop" (i.e., likely burn) decision class; any blue pixels seen are representative of the "NBR increase" (i.e., possible regrowth) decision class. The final BULC-D probabilities layer containing all three of these decision classes was overlayed with the Confident Burn layer, displayed in pink, which filtered for only those pixels that were most likely to have burned since the expectation year. The conditions to satisfy for pixels to be selected for this filter are different for NBR and BAI. For NBR, a confident burn is any pixel where the mean NBR value in the expectation collection (year i) is greater than 0.4 and the BULC-D probability of the NBR drop decision class in the target collection (year i + 1) pixel is greater than 50%. For BAI, a confident burn is any pixel where the mean BAI value in the expectation collection pixel is greater than 50 and the BULC-D probability of the BAI increase decision class (i.e., the probability of having burned) in the target collection pixel is greater than 50%.

1) 2020 Cameron Peak Fire, Colorado

    The final BULC-D probabilities output for both BAI and NBR were able to distinguish true fire pixels from non-fire pixels within the observation period between day 135 and day 290. A large swath of the Eastern part of the fire, which was shown to have burned in the official extent map, was not shown to be burned in neither the BAI nor the NBR output. It is important to mention, though, that a large snowstorm swept through the region on October 24th (day 297) and 25th (day 298). It was found that, while limiting the observation period to just before this snowstorm prevented any analysis of the eastern part of the fire, it proved important for maintaining integrity of the other parts of the burn scar. Once snow fell in the region, it blanketed the entire scar, thus negatively impacting any interpretations of the final output after snow fell. When the observation period was extended beyond day 300, the snow storm caused BAI values to fall and NBR values to rise. As a result, fire behavior dramatically decreased and BULC-D was no longer that true burn pixels had actually burned. While the effect of this snowstorm on fire behavior is an interesting and important finding for better understanding wildfire dynamics as a whole, the snowstorm was excluded from the observation period. Before this project began, it was determined that, even without considering the secondary effect of non-fire phenomena, NBR did not perform well in the montane, semi-arid environment where the Cameron Peak fire took place. As such, including the snowstorm in the observation period did not serve this project's goal of determining how BAI-based burn scar visualization performs in regions where NBR-based burn scar visualization has been shown to not performed well. 

2) 2020 East Troublesome Fire, Colorado

    BULC-D was largely unable to detect or visualize burn scar activity in the East Troublesome fire region, regardless of whether it used NBR or BAI. Interestingly, neither BAI nor NBR were able to distinguish true fire pixels from non-fire pixels within the observation period between day 170 to 325. It is hypothesized that this is due to the region being affected by the same snowstorm that was excluded from the observation period in the Cameron Peak test fire that occurred on October 24th (day 297) and 25th (day 298). One could argue that because the snowstorm did not start until day 297, there are still plenty of images available between day 170 and then. However, the fire did not officially start until October 14th (day 287), only amounting to roughly 10 days-worth of images that may contain true fire pixels. As a result, it seems that the period between when the fire actually started and when the snowstorm began may have been too short to accrue more than one or two images containing true fire activity, which may explain why BULC-D was unable to capture this fire regardless of the index it used.

3) 2020 Chernobyl Exclusion Zone Wildfire, Ukraine

Neither NBR-informed nor BAI-informed BULC-D outputs were able to detect or visualize true fire pixels present in the region according to the official fire extent map. Burnt regions identified in the official extent map did not match with those identified in  either output; it also seems that the two indices rarely agree with one another, as there is little to no discernible pattern that is equally emphasized in both outputs. The final BULC-D probabilities output developed with both indices were so different, yet a high proportion of all burn decision classes identified were also colored pink and identified as being confident burns. This fire was originally chosen because the Red Forest, composed mostly of coniferous trees, was expected to be somewhat similar to the ecology of forests where the Cameron Peak and East Troublesome fires took place. The region is dissimilar from the Colorado fires, however, in its elevation profile and topography. As well, the near complete dissimilarity in BAI and NBR's performance in this region is also a unique difference between BULC-D outputs in this region as compared to how the indices performed in the Cameron Peak and East Troublesome fire regions. 

4) 2019-2020 Kangaroo Island Bushfire, Australia

The Kangaroo Island bushfire is the only of the test fires to have occurred over two separate years, spanning from December 20th, 2019 to January 21st, 2020. Because of this and the fact that BULC-D distinguishes each collection of images it gathers from Landsat 8 and Sentinel 2 by year, the expectation and target collections being provided to BULC-D are inherently incomplete. Despite this, both NBR and BAI perform well in comparison to the official fire progression map, though inter-index variability is noticeable when comparing the two outputs. Burned pixels identified in the final BULC-D probabilities outputs of both indice  almost completely match confident burn pixels, suggesting that the pre and post-fire signatures of both indices varied significantly. Some decision classes in the BAI output image disagree with NBR decision classes in the same region. For instance, a small grouping of pixels in the southwest corner of the island was classed as "BAI decrease" (i.e., no burn or regrowth) by the BAI-informed output, but was confidently burned in the NBR output. The official fire extent map seems to align more closely with the NBR output image, suggesting that BAI values may have dropped significantly after rising at the time of the fire, altering the final decision class that BULC-D assigned to these pixels. It was initially hypothesized that this was somehow due to the incident period being divided between 2019 and 2020, and that the fire signature within regions that had burned in the 2019 was not being captured or recognized when the algorithm was using 2020 as its target year. This is likely not the case, however, as the observation period creates the expectation collection and the target collection from images collected during the same part of the year. This means that the expectation collection spans day 1 to day 77 of 2019, and the target collection spans day 1 to day 77 of 2020. Thus, the only impact that this division of the incident period has on fire detection and visualization is that fires that burned in 2019 may no longer be burning by the beginning of 2020. This could explain the "BAI Increase" signature that the BAI output picks up on in the southwest corner of the island, as fires that burned in the western half of the island during 2019 could have peaked and fallen by the end of the year. 

Challenges & Limitations

Challenge 1: Learning how BULC-D works

• Modifying BULC-D to be able to use BAI proved relatively complex and time consuming. More time consuming, however, was the process of understanding and explaining how the algorithm actually works. To mitigate this limitation and stick to deadlines while also maintaining project objectives, I conducted a lot of research to better inform myself on the theories and methods behind how the algorithm works. I believe this supplementary research allowed me to better disseminate my own knowledge and explain the complexities of this project. 

Challenge 2: Implementing the BAI conversion factor

• • There was not much information available online regarding developing a conversion factor for BAI when using Surface Reflectance imagery. As a result, the process of implementing a conversion factor for the BAI equation required trial and error testing. While it is inevitable that NBR values, lying between 0 and 1, and BAI values, ranging anywhere from 1 to above 1000, will not lie perfectly in the same range, it was important to get these values as close to one another as possible so that the range of possible BAI values would ideally lie within the same relative order of magnitude as NBR values. Moreover, an ideal conversion factor would be one that has BAI spanning its native value range when using Top of Atmophere imagery. Thus, I went around to different fires looking BAI values within individual pixels in an SR image before and after implementing a candidate conversion factor and determined whether or not the value was satisfactorily similar to the BAI value at that pixel in the TOA image. This was a very roundabout way of finding the conversion factor, however, as TOA to SR calibration is standardized with a conversion factor of 10000, though I did not figure this out right away because I was implementing the conversion factor in the wrong segment of the BAI equation. My work was complete when I found that the correct placement was to have each of the spectral bands, Red and NIR, be divided by 10000.

Challenge 3: Tweaking BAI's custom transition matrix

• • While collecting samples of BAI values across the two fires I looked at was relatively straight forward, the process of tweaking and adjusting the new index's custom transition matrix required a significant amount of trial and error. The histograms I developed helped to gain a better understanding of BAI's overall distribution, but much of the work of ensuring that decision classes were being accurately conveyed in BULC-D outputs on the map was quite time consuming. This was just part of the deal, however, as not doing so would have negatively impacted the algorithm's ability to accurately and reliable visualize a given fire. 

Limitation 1: Manual verification via point analysis still required

• Manual verification via point analysis was still needed to come to firm conclusions regarding burn status of a pixel. This meant clicking around on individual pixels at each of the four test fires and coming to my own judgements about whether the decision class it was placed within was the correct one based on the overall distribution of that index throughout the expectation and target year. While this may sound like more of a challenge, it can easily become a limitation if the decision classes that index values are being placed in are inaccurate and unrepresentative of the true state of the land cover on the surface. It may be instinctual to trust that BULC-D is doing a good job when you see how it has performed in the Cameron Peak or Kangaroo Island fires, but one must always be aware of the confounding effect of non-fire phenomena, like snow or smoke, and incompatibility of a given index for certain regions or ecologies. 

Limitation 2: Some test fires may not have been suitable candidates for comparison

• There is a chance that the candidate test fires I have chosen will not be suitable for this project, or that conclusions or relationships drawn between the fires – particularly those between the Colorado and Ukraine fires – will be less meaningful than I anticipate. If I come to this realization early on in the project, I may have time to swap the Ukraine fire for a more suitable alternative. Otherwise, I will simply note any lack of meaning between the two as a limitation to the project.

Conclusion

This project aimed to repurpose the BULC-D algorithm to utilize the Burned Area Index (BAI), with the ultimate goal of developing a new "spectral lens" from which wildfire burn scars can be analyzed. Two candidate test fires – the 2020 Cameron Peak and East Troublesome fires in Colorado – were chosen based on the disagreement between the fire visualizations created by NBR-informed BULC-D and by corresponding fire progression/extent maps. A third candidate fire – the 2020 Chernobyl Exclusion Zone fire in Ukraine – was chosen for the relative similarity of its environmental conditions to those of the Colorado fires. Lastly, a fourth control fire – the 2019-2020 Kangaroo Island bush fire in Australia – was chosen for the relative similarity of the region's environmental conditions to the arid/semi-arid regions where the NBR-informed BULC-D has historically agreed with or improved upon official fire progression maps. It is from these four fires that the project aimed to answer the following questions:

• How well does a BAI-informed BULC-D visualize burn scars as opposed to official fire progression or extent maps of the four test fires?

• How does BAI-informed BULC-D perform compared to its NBR-informed counterpart in:

  • • regions where NBR-based burn scar visualization has performed well?

    • regions where NBR-based burn scar visualization has not performed well?

The central aim of this project was accomplished, with BULC-D now capable of detecting, visualizing, and analyzing fires with BAI. Cross-comparison between NBR and BAI BULC-D outputs yielded varying levels of compatibility in the four test fire regions. Perhaps most interesting is the opportunity posed by inter-index comparison and the potential to combine BAI and NBR outputs into one contiguous, multi-index informed output. The resulting algorithm would emphasize decision classes that observed inter-index agreement between BAI and NBR, as well as dampen decision classes that observed inter-index disagreement. Overall, this would ideally increase the accuracy, flexibility, and usability of the BULC-D algorithm with respect to post-fire burn scar detection and visualization. 

Methodologies that were carried out in the project were outlined, as were potential conceptual and practical problems that were encountered while striving to complete the project. Deliverables in the form of comparative maps, script excerpts, and explanatory graphs were delivered.

References and Resources

(1) United Nations Environment Programme. (2022, March 9). As climate changes, world grapples with a wildfire crisis. UNEP. Retrieved March 11, 2022, from https://www.unep.org/news-and-stories/story/climate-changes-world-grapples-wildfire-crisis

(2) Willis, E. W., & Cardille, J. A. (2021). Applying Harmonics to Satellite Burn Indices for Wildfire Burnscar Mapping: Bayesian Updating of Land-Cover (BULC-D). [Conference presentation]. Agricultural and Environmental Sciences (AES) 2021 Undergraduate Student Research Award (USRA) Poster Presentation Event. Conference online due to COVID-19. https://drive.google.com/file/d/1t_nqV7r9JyUCPq0TgHgiBeNyGzOI7v1f/view?usp=sharing

(3) Cardille, J. A., & Fortin, J. A. (2016). Bayesian updating of land-cover estimates in a data-rich environment. Remote Sensing of Environment, 186, 234–249. https://doi.org/10.1016/j.rse.2016.08.021

(4) Key, C. and N. Benson, N. "Landscape Assessment: Remote Sensing of Severity, the Normalized Burn Ratio; and Ground Measure of Severity, the Composite Burn Index." In FIREMON: Fire Effects Monitoring and Inventory System, RMRS-GTR, Ogden, UT: USDA Forest Service, Rocky Mountain Research Station (2005).

(5) Chuvieco, E., Martín, M. P., & Palacios, A. (2002). Assessment of different spectral indices in the red-near-infrared spectral domain for burned land discrimination. International Journal of Remote Sensing, 23(23), 5103–5110. https://doi.org/10.1080/01431160210153129 

(6) USDA Forest Service, Fire and Aviation Management. (2021, June 21). Cameron Peak Fire Information. InciWeb – The Incident Information Service. Retrieved March 22, 2022, from https://inciweb.nwcg.gov/incident/6964/

(7) USDA Forest Service, Fire and Aviation Management. (2021, June 21). East Troublesome Fire. InciWeb – The Incident Information Service. Retrieved March 22, 2022, from https://inciweb.nwcg.gov/incident/7242/ 

(8) Wikipedia contributors. (2021, December 15). 2020 Chernobyl Exclusion Zone wildfires. Wikipedia. Retrieved March 22, 2022, from https://en.wikipedia.org/wiki/2020_Chernobyl_Exclusion_Zone_wildfires

(9) Local Recovery Team, Kangaroo Island Local Recovery Committee (2020, November). Kangaroo Island Community Recovery Plan 2020 – 2022. Government of South Australia. https://www.recovery.sa.gov.au/2019-20-bushfires/kangaroo-island/kangaroo-island-icon-panel/our-community/our-community/KI-CommunityRecovery-Plan-Final-LoRes.pdf 

(10) ESRI. (2022). Indices gallery—ArcGIS Pro | Documentation. ArcGIS Pro. Retrieved March 22, 2022, from https://pro.arcgis.com/en/pro-app/latest/help/data/imagery/indices-gallery.htm

(11) Landsat NASA. (2022, February 16). Landsat Science. Landsat Science | A Joint NASA/USGS Earth Observation Program. Retrieved April 17, 2022, from https://landsat.gsfc.nasa.gov/

(12) How to calculate Z-scores (formula review) (article). (n.d.). Khan Academy. Retrieved April 17, 2022, from https://www.khanacademy.org/math/statistics-probability/modeling-distributions-of-data/z-scores/a/z-scores-review 

(13) Ord, K., & Stuart, A. (2010). Kendall’s Advanced Theory of Statistics, Distribution Theory (6th ed., Vol. 1). Wiley. https://openlibrary.org/books/OL1196159M/Kendall’s_advanced_theory_of_statistics.

Initial proposal

Here's a link to the proposal of this project.