The purpose of this study is to make a quantitative comparison of Fire Potential Index (fpi) rasters created by identical processes and on coincident dates, differing only in the sensor from which the source data were obtained.
Historically, fpi rasters have been produced from Advanced Very High Resolution Radiometer (AVHRR) data, with the latest such instrument aboard NOAA-15. More recently, fpi has been derived from similar data obtained from the Moderate Resolution Imaging Spectrometer (MODIS) aboard the Aqua satellite. The motivation of this study is to determine what if any accommodation needs be made to use fpi data derived from the latter source.
The fpi comparisons were performed on datasets from several hundred dates for which both AVHRR- and MODIS-derived products were available. Specifically, 301 dates between 4/3/2017 and 2/14/2018, and an additional 58 dates between 7/4/2018 and 9/1/2018. All rasters were pre-processed by selecting only those pixels that reflect computed fire potential, discarding other background values such as those indicating water bodies.
The initial metric considered was the simple arithmetic difference between two rasters of the same date, always computed as fpiMODIS - fpiAVHRR, where the range of possible values for fpi range from 0 to 100 in both cases. The resulting rasters were visually inspected as a first step, to confirm good spatial co-registration. No issues were identified. Summary statistics of Mean and Standard Deviation were noted for each difference raster and were stored in a spreadsheet. Additionally, full histograms (count of each difference vaklue – all integers) were extracted and stored.
The difference rasters reveal that the MODIS-based values are generally higher than those based on AVHRR. Over the 359 dates considered, values range from +1.63 to -0.07, with an average value of 0.94. Typical standard deviations are correspondingly small, with an average value of 0.92.
An examination of figure 1 reveals fairly well-defined regional variations in potential index differences, which will now be examined in some detail. To limit consideration to systematic differences, a new raster was computed as the average of 300 difference rasters from the first group (4/3/2017 and 2/14/2018). Some regional differences persist, as shown in Figure 2.
The average value of pixels in the 300-difference raster is 0.92, but with visible spatial variations. The above raster could be divided in any number of ways, but one that may be of particular relevance is the management divisions of the Geographic Area Coordination Centers (GACC) established by the National Interagency Fire Center (NIFC).
The 300-difference raster was subdivided into the nine GACC areas of the contiguous U.S., and statistics were noted for the pixels in each area, as shown in Figure 3.
In addition to revealing a spatial component to fpi differences, many of the histograms in Figure 3 betray a positive skew, suggesting an underlying systematic bias. Reviewing the data in detail suggested that the MODIS-based data were larger in magnitude by a nearly constant factor, rather than by a fixed offset.
All of the 359 source fpi rasters were reprocessed to obtain the ratio fpiMODIS / fpiAVHRR, which was found to have a value of roughly 1.07 both within and between the various dates. To subtract the effect of this factor, all AVHRR-based fpi rasters were multiplied by their average fpiMODIS /fpiAVHRR factor, and differences were again computed, now based on the formula fpiMODIS - fpiAVHRR_ADJUSTED.
Repeating the previous analyses on the adjusted differences revealed that the histograms were more normal in appearance (example in Figure 4), and, more significantly, that the mean value of the 300-difference rasters was reduced to a negligible 0.04.
A visual inspection of the adjusted 300-difference raster reveals a strongly similar appearance to unadjusted difference raster shown in Figure 2. It is thus concluded that some persistent geographic biases may be attributed to differences in the respective sensors. These differences are too small in magnitude to warrant further investigation.
A detailed review of Fire Potential rasters produced from both AVHRR and MODIS data shows them to be nearly identical for practical purposes, with MODIS-derived rasters exhibiting a roughly 7% greater value than those derived from AVHRR, and with typical absolute differences on the order of 1 (in the percentage units of the potential rasters). Some persistent geographic disparities are observed even when other systematic biases are taken into account, but are small enough in magnitude that they need not be considered in application of the data.
Download the full study report: A Comparison of Fire Potential Index Rasters Derived from AVHRR and MODIS Data