Global fire activity is changing in many areas as temperatures increase and land use intensifies1,2,3,4,5. This is sparking an increase in attention given to fire activity and fire ecology. However, the availability of data for spatially delineated fire events is limited or non-existent in many countries6, with most global fire data coming from satellite-based active fire detections7,8 and gridded burned area products9,10. The lack of products containing delineated events has led to many global studies about fire ecology that are computationally-intensive, coarse-scale trend analyses1,4.
A key advantage of datasets like Monitoring Trends in Burn Severity (MTBS)11 or the Fire Occurrence Dataset12 lies in their ease of use. Since its inception in 2007 MTBS has been cited 947 times in peer-reviewed studies according to a Google Scholar search at the time of this writing, despite documented limitations for scientific use of some facets of the product13. The MTBS dataset is regularly updated, easy to find on the internet, and it is free, fast and easy to download and use. Many environmental scientists and resource managers do not have the computational budget or expertise in big data or remote sensing to deal with the challenges one must overcome to process large fire datasets. This is especially true for cases when all that is needed is a shapefile of fire perimeters that can be used to map fire history. Other global fire perimeter datasets have been produced from satellite-derived burned area products14,15, but these are only available in yearly or monthly global shapefiles. Often field-based studies of fire effects require an entire time series over study areas that are only a few hundred km in diameter16 or a single ecoregion17. The end user who wants to understand the fire history for their region would have to download yearly shapefiles with a global extent, clip all of those shapefiles to their area of interest, and then combine them into one shapefile, just to get started. We suspect that the lack of accessible fire perimeter datasets that are easy to download and use contributes to a disparity in research, where fire ecology studies are conducted mostly in developed countries that have either research infrastructure capable of handling big data or longer-term government records, or temperate forested regions that have substantial tree-ring records18.
There are two existing global perimeter products, the Global Fire Atlas (GFA) (Andela et al.14) and the Global Wildfire Information System (GWIS) (Artes et al.15). Both were created by applying spatiotemporal flooding algorithms to the MODIS MCD64 Burned Area Product. These algorithms assign burned pixels from the MCD64 products using a moving window whose size is defined by spatial and temporal parameters. They are created as monthly or yearly slices of the entire globe, and they can be subsetted. These products are extremely valuable for global scale studies. But when we look at how those products delineate known fire events we see a consistent problem in that they both seem to over-segment events in ways that appear unrealistic. This inconsistent event delineation is not problematic for coarse-scale or regional estimates of burned area or fire seasonality, but can lead to unrealistic estimates for number of fire events and event-level characteristics like fire size and spread rate. In Fig. 1 we illustrate this with an example of the 2013 Rim Fire in California, United States, which was unmistakably a single event that burned about 90,000 ha over the course of three months. Figure 2 illustrates how the day-to-day progression of the Rim Fire was a steady progression from a single ignition in late August. Table 1 shows how the differences in event delineation propagate to calculations of burned area and number of events. In the GFA, the Rim Fire is delineated as one large event of 804.5 km2, and 13 additional events totaling 88.7 km2. in GWIS it is delineated as one event of 878 km2 and 47 additional events totalling 20 km2. With FIRED, there is one event of 892 km2 and 2 single pixel events totalling less than one km2. One cause for potential differences is how one defines a “fire event”. Large fires often have multiple ignition sources. The Global Fire Atlas algorithm and others19, for example, search for local minima to identify various ignition locations that may begin as small patches, only to later form a large complex and in the end described with a single fire perimeter. The choice of outside sources for optimizing the spatial-temporal parameters, the method of optimization, and the intent of the final product’s meaning (defining events as single ignition patches vs contiguous burned area) all lead to different outcomes in the final events that are delineated. Another likely source of this discrepancy is that GWIS and GFA are calibrated to create a single global product. Because different geographical areas have different types of fire regimes, they have fires that grow at different rates and to different sizes, and occur in greater or fewer frequencies, and so the spatial and temporal parameters that work well for defining a fire event in one area may result in over- or under-segmentation in other areas. Here, we decided upon an approach of creating many regional products across the globe, rather than one product for everywhere on earth.
Comparison of global fire event products performance for the 2013 Rim Fire (a). In the FIRED product (b), the Rim fire was classified as one very large event with two single pixel events. The Global Fire Atlas (GFA, c) and Global Wildfire Information System (GWIS, d) each delineated a very large event, with 13 and 47 smaller events, respectively.
The two primary outputs FIREDpy provides are a daily- and event-level product. Panel a shows the default single event polygon. In b, each day has a separate polygon, with associated statistics generated, within each event. Panel c shows the daily perimeters derived from the airborne infrared by the incident management team for comparison.
Besides the ease of access and use, the advantage of the FIRED product lies in the user’s ability to use the open-source software, FIREDpy, to tailor the spatial and temporal parameters of the moving window algorithm in order to realistically delineate events for their region of interest. In Fig. 3, we illustrate this by comparing the three products for a pair of small fires in Florida. In this case, the FIRED product that was created with a larger moving window (5 pixels and 11 days) over-aggregated the events, but it only required one line of code at command line to recreate the product with a smaller moving window (1 pixel and 5 days) to get more realistic results.
Product comparison for two small events in Florida, the Moonshine Bay and Sour Orange fires (outlined) that both ignited in February of 2007 and were delineated by MTBS. In b the firedpy product that was optimized for the entire United States with a moving window of 5 pixels, 11 days resulted in aggregation of the two fires delineated by MTBS, but also several smaller fires nearby. In b, it was re-ran with a window of one pixel and five days, for a more realistic result. Delineations by the Global Fire Atlas (c) and the Global Wildfire Information System (d) are shown for comparison.
Here, we present a collection of regionally-tailored fire perimeter datasets for every country in the world with significant fire activity20, which we created with the open source algorithm, FIREDpy21. Each dataset is either a single country or a broader region, depending on the data volume. These datasets differ from other similar efforts14,15 in that each dataset created by FIREDpy is a single file containing a collection of polygons that is generated for the entire time series, rather than monthly or yearly aggregations with a global extent. Furthermore, we have generated the data products at a spatial extent land managers and ecologists would typically use to do regional-scale research, and we adjusted the spatial and temporal parameters for each country to yield realistic event delineations. We also made every effort to ensure that download sizes are reasonable ( < 300 MB). They have a temporal extent from November 2000 to the summer of 2021 at the time of this writing, and they will be actively maintained and updated yearly and upon request. Most importantly the software we developed to generate the datasets is open source and freely available, and so novel fire perimeter datasets can be generated by anyone for any area of interest at any time, using a single command. We hope this will increase the capacity of ecologists and land managers across the world to study fire activity, and incorporate fire history into their work. We invite the broader community to contribute to the continued development of the software package and associated data products.
Source: Ecology - nature.com
