Remote Sensing as a Tool in Wildfire Inventory Creation

 

By Ian Mullet

December 4th, 2004

 

Introduction

            During the August-September 2000 period an extensive air quality study was conducted throughout Texas, entitled the Texas Air Quality Study (TexAQS). The data collected has since been used to test the outputs of air quality computer models, in the hope of having these models serve as tools to verify that industry meets air quality standards set by the Environmental Protection Agency (EPA) and the Texas Commission for Environmental Quality (TCEQ). To test the models properly, a comprehensive set of model inputs need to be produced.

A database of wildfires during this modeling period serves as one input. By coincidence the same period chosen to conduct this large study, also happened to be a period of intense wildfire activity. However, the creation of a wildfire inventory has proved quite tricky, for two reasons. First, it has been a labor intensive task. Several different independently created records of wildfires have to be tracked down from numerous sources. Second, these records prove surprisingly inaccurate at times. While it would seem simple to record the location of a large fire, the current sources sometimes locate large fires in entirely wrong counties. A solution must be found.

Remote sensing of fires serves as a potential remedy to both these difficulties. Remote sensing, in this case, means the use of instrumentation on board current satellites to detect the location and potentially the intensity of wildfires. The MODIS (Moderate Resolution Imaging Spectroradiometer) instrument, currently located on board three polar-orbiting satellites (though only onboard one in the year 2000), measures the intensity of radiation entering its chamber at 31 different bands, monitoring wavelengths from the visible spectrum to infrared. Furthermore, an algorithm has been developed to detect thermal anomalies (e.g., fires or volcanoes) occurring in the swath of the satellite. The basic concept behind this algorithm is that a large change in the typical radiation for a given location is flagged as a fire. The output from this algorithm is called in this paper the fire-mask. Use of this fire-mask can potentially replace and certainly assist the current methods for creating a wildfire inventory. The questions become: a) can use of an automated fire-mask serve as a single source for a wildfire database, thus reducing the labor necessary in creating a fire inventory for future modeling periods and b) can we rely on it to detect accurately the fires in a given area?

            To answer yes to the first question we must be able to answer yes to the second. To accurately detect fires, the fire-mask must not produce false positives (state a fire exists where one in fact doesn’t) nor produce false negatives (miss a fire).

            Therefore, this paper will compare the fire-mask to the current inventory manually culled from disparate sources in the hopes of correcting as best as possible mistakes in the original. Also it will explore the utility of using remote sensing as the primary tool in the creation of databases for future modeling periods.

 

Methodology

 

            The general strategy to achieve the above goal is as follows:

a)      load the manual inventory into GIS

b)      load the results from the fire mask,

c)      compare the two to gain a general sense of how well they correspond,

d)      choose select days and regions to compare the two different inventories to all visual and thermal images available from satellites to verify fire locations.  

 

The manual inventory was created through the long work of Victoria Junquera, a master’s student in the research group of Dr. David Allen in the chemical engineering department. It’s creation involved culling together data from several different sources of varying exactness. Sometimes size of fire was not reported, only the tally of fires in a county or region for a given month. Interpolation methods were often used to estimate the size location and time of fire. Sometimes the location given was far off. While it had many inherent errors, however, we must also point out that it was very comprehensive.

 The data for the manual inventory was stored in an excel spreadsheet with columns for date, fire size in acres, latitude and longitude, along with numerous other features not pertinent to this study.

This excel worksheet was ultimately displayed as a layer following these steps:

a)      Convert .xls file to dbase IV file in Excel called “mandata” as in manual data.

b)      Create a new personal geodatabase called “wildfires”.

Figure 1-3 Screen captures of the process described

c)      Create a feature class from XY table of “mandata” and project it in NAD 1983.

 

d)      Export “xymandata” into personal geodatabase. Now a shapefile is in “wildfires” geodatabase.

e)       

 

f)        Open ArcMap and load data as layer.

Figure 4 A display of map with data as it first appears. MODIS in green, manual fires in red. What a mess!

 

The MODIS data was delivered to me by Solar Smith, a research scientist at the Center for Space Research, here at the University of Texas, in a CSV file. Which can easily by displayed in Excel. After collecting the data from all of the dates into one file and trimming the fires the were outside of the area of interest, I saved the file as an Excel file and  followed the same steps as above to display them as a layer. Above is a screen capture of the two datasets along with a map of Texas for reference.

            The data displayed in this manner reveals very little. One must adjust the properties of these two layers to gain any insight from these data. The three most important property adjustments are: 1) display by date 2) display manual fires only above a certain size threshold 3) show a size gradient of manual fires

            The “Definition Query” tab under “Properties” accomplishes the first to goals. It allows one to display the fires by date so that only fires on, say, September 5th are displayed. A second important filter created by “Definition Query” is the elimination of fires under a certain acreage in the manual database. We should not expect MODIS to detect a one-acre fire that is recorded in the manual database. The size threshold for this project has been set at 50 acres.

            The settings on the “Symbology” tab can be set to display the data so that a size gradient for the manual fires will be shown. This allows one to see at approximately what size a fire must be before MODIS will pick it up. Also, it brings to attention any reportedly large fires that do not have a MODIS detection near it.

            Once these daily results have been visually inspected and had conclusions drawn from them, the next task is to verify the fire locations by using available satellite imagery. This task is achieved by loading .bsq files of the imagery into ArcMap. ArcMap “help” describes a binary sequential file (BSQ) as “an ASCII description file that describes the layout of the image data. Black-and-white, grayscale, pseudocolor, and multiband image data can be displayed.” These files are very large, however, and images of the complete domain for all days is not a viable option. Instead, imagery in visible and thermal bands were gathered in the southest region of Texas for the days of September 5th and 6th, days on which large wildfires occurred in Texas and Louisiana, and days on which the meteorological conditions were such that smoke plumes could be more easily detected. These images were overlaid with the locations of the fires in both databases so that immediate comparison could be made. Should a smoke plume be seen near a MODIS fire then one can conclude that MODIS correctly flagged that location. This fire should be included in any final database. Likewise, an apparent fire with no MODIS location nearby will show that MODIS is capable of missing fires and thus test the accuracy of the fire-mask. Likewise the manual database can be tested. Should MODIS be regularly confirmed then it becomes safe to relocate fire locations in the imprecise manual database to those indicated by the MODIS fire-mask. Different scalings for color representation are experimented with in an effort to best bring out smoke plumes, thermal anomalies or any other indications of a fire presence. Furthermore, bands are scanned to see which one was best able to capture the fire activity (e.g., do visible bands spotlight the thermal acticity better than thermal bands for this particular image?)

           

Results

 

            Prior to confirming fire locations with thermal imagery, daily maps of the two fire databases were created, with manual fires only above a fifty acre threshold appearing. The results were at times disconcerting, seeming quite chaotic. But certain patterns did emerge. 

            One might hope after placing these settings on the data a tidy display would result. Certainly there would be differences between the two, but most large fires would have a MODIS fire somewhere near it and vice versa. It was my initial hope that I could visually connect any major manual fire to a proximate MODIS fire, and thereafter adjust the coordinates in the manual database, to the more accurate ones generated from the fire-mask. However, as the images in the appendix show, such a simple plan could not come to fruition. Many MODIS fires did not have a nearby manual fire and vise versa. If we look at the simple day-by-day results, however, certain patterns do emerge. As the manual fires become larger and larger the likelihood of a proximate MODIS fire increases. On the other hand the likelihood of a MODIS fire being unaccompanied by a manual fire increases dramatically as one leaves the state of Texas and moves east.

            Below is an from MODIS from September 5th, 2000. Nine MODIS fire indications (represented by black triangles) exist within the bounds of the particular image. These indications are clustered into five general locations. For three of these one can immediately detect the presence of fires because of the large smoke plumes generated. For the other two locations no plume could be detected. Nor could a plume be detected for the 6000 acre fire recorded for that day in the manual database. For the three fires with plumes no manual fire was in the immediate area, the closest lacation being one-third of a degree off.

 

Figure 5 MODIS image in mid-infrared band of September 5th, 2000, used to confirm fire locations.             

 

            September 6th was another day of extensive fire activity and long plumes, ideal for visual confirmation of the fire inventories. On this day, the image was most revealing when viewing a band from the visible portion of the spectrum. Once again the major plumes are captured by the MODIS fire-mask. The manual database also does well with fires reported in the general area of several major plumes. In this image, we see MODIS not capturing the plume that is visible in the very center of the image. At the same time, we see the manual fire database not registering fires in the southeast, northeast, and south central. Many of the indicated fires are difficult to visually confirm.

 

Figure 6 MODIS image from the visual spectrum band on September 6th, 2000.

 

 

 

Conclusions

 

The dramatic drop-off in manual fires outside of Texas indicates that either records were lacking for wildfires in Louisiana or not as much effort was put into gathering those records and inputting them into the manual fire database. It must be decided whether fires from this source affect the modeling domain of Texas enough to warrant making efforts to include them in the database. The general lack of MODIS fire indications near fires less than 500 acres shows that this may be an approximate size threshold for the fire mask.

            Visual confirmation of many of the MODIS fire locations demonstrates that the MODIS fire-mask is a valuable tool for registering more exact locations of wildfires. Unfortunately, it is difficult to detect the presence of a fire without a plume, leaving the accuracy of both tools still in doubt when no plume is present. As an attempted remedy, future work will include the inspection of imagery from other selected satellites, whose data can be obtained from the Center for Space Research. This includes two NOAA satellites orbiting at the time along with another satellite named Orbview2.

            Quantitative analyses also seem appropriate future work. These may include quantifying the number of fire events with matches in both databases, comparing the size of fires for matches (to gain an idea of the MODIS threshold), determining the number of fires whose location should be corrected, and calculating the total length of those corrections. Ultimately, the test of how important it is to correct the database will come not through GIS analysis but analysis of the model simulations that occur with the new database serving as input.

 

 

Sources

The MODIS fire-mask and satellite images were delivered to me by Solar Smith of the Center for Space Research, University of Texas at Austin.

The manual database was compiled by Victoria Junquera of the University of Texas at Austin.

 

Appendix

Below are selected maps from ArcGis of the daily fires. The triangles represent fires indicated by the MODIS fire-mask. The circles represent fires from the manual database.