Fire Detection Misalignments Between GOES ABI and VIIRS and Their Impact on GOES FDC Evaluation

Highlights

What are the main findings?

• Spatial misalignments between GOES FDC and VIIRS detections occur in approximately 12% of fire events across diverse latitudes and ecosystems.
• Implementing a buffer significantly mitigates the misalignments impact, reducing estimated false alarm rates from 26–36% down to 7–15%.

What are the implications of the main findings?

• Standard accuracy evaluations that treat VIIRS as ground truth without spatial buffering can yield biased and unreliable estimates.
• The proposed evaluation scheme is generalizable and can be applied to assess other geostationary and Low Earth Orbit sensor combinations.

Abstract

Wildfires cause major damage, and their accurate detection is crucial. A common approach to near-real-time detection uses Geostationary (GEO) satellite algorithms. A standard scheme for evaluating the accuracy of a GEO-based algorithm is to compare its detections with higher-resolution Low Earth Orbit (LEO) images, considering the latter as ground truth. The primary objective of this study is to quantify the prevalence of GOES ABI/VIIRS fire detection misalignments and assess their impact on the accuracy evaluation of the GOES Fire Detection and Characterization (FDC) product. Thus, the key question is how this evaluation should be performed. To this end, a large dataset of matching FDC/VIIRS fire detections across Western U.S., Amazonas, and Patagonia was constructed. Our finding is that for nearly 12% of fire events, there are spatial misalignments between FDC and VIIRS detections. Next, we show that using VIIRS as ground truth without considering these misalignments yields highly biased estimates. This affects the evaluation of the FDC product detection capabilities. Finally, we demonstrate that using a GOES FDC/VIIRS buffer window substantially mitigates the effect of misalignments. For example, the estimated false alarm rate ranges between 26% and 36% without a window, whereas using a $3\times 3$ window yields values between 7% and 15%.

1. Introduction

Climate change, with its severe heat waves, prolonged droughts, forceful wind gusts, and changes in precipitation patterns, is causing more frequent and intense wildfires [1]. For example, 13 of the largest wildfires on record in California occurred within the last ten years, with nine of them in 2020 and 2021 alone [2]. Wildfires cause major disturbances to natural ecosystems and human communities, including loss of human lives [3]. Fire detection and monitoring are thus crucial components to mitigate wildfire damage.

One approach to detect and monitor fires is to use images from Geostationary Earth Orbit (GEO) satellites [4]. This approach can detect fires in near real time since GEO satellites capture images every 5–15 min. Indeed, fire detection products have been developed for most GEO satellite systems, mainly using threshold-based algorithms [5,6]. Notable examples of GEO satellites and sensors used for fire detection are the Advanced Baseline Imager (ABI) onboard the Geostationary Operational Environmental Satellite–R Series (GOES-R) [7], the Himawari-8/9 that is equipped with the Advanced Himawari Imager (AHI) [8], and the Geo-Kompsat-2A carrying the Advanced Meteorological Imager (AMI) [9]. Over the past few years, several efforts have been made to enhance the fire detection capabilities of current GEO-based algorithms. In particular, several authors developed sophisticated machine learning-based fire detection algorithms [10,11,12]. Hence, evaluating the performance of a given fire detection algorithm is a key component, both for comparing the accuracy of different methods and for developing improved ones.

The omission and commission error rates are two common quantities used to evaluate fire detection algorithms [13]. The omission error rate is defined as the probability that the algorithm does not detect a fire. The commission error rate, also known as the false alarm rate, is the probability that the algorithm declares a fire even though there is no fire at that location [13]. Other related evaluation measures are the precision and recall of a fire detection algorithm, and its corresponding F1-score [14].

To estimate these quantities, it is essential to have a reliable ground truth reference for comparison. One of the most accurate approaches is to use ground-based in situ fire reports as a reference, as collected by disaster response entities such as a National Forest Service [12]. However, such reports may have limited spatial coverage and are often not publicly available. A different approach is to use the fire product output from images of Low Earth Orbit (LEO) satellites as ground truth. LEO satellites have higher spatial resolution than GEO satellites, and their fire products are considered much more accurate [15,16]. This approach has been applied with several LEO satellites and their imaging systems, including the Visible Infrared Imaging Radiometer Suite (VIIRS), Moderate Resolution Imaging Spectroradiometer (MODIS), Landsat, and Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) [13,17,18,19,20,21,22,23,24,25].

However, as noted in several studies [26,27], there can be spatial misalignments in the locations of objects as they appear in LEO and GEO images. The fundamental drivers of these GEO-LEO spatial misalignments are related to physical and geometric constraints. A primary factor is the parallax effect. GEO sensors are located over the equator and often observe regions at large zenith angles. In contrast, LEO sensors might view the same area from a near-nadir position or from a large viewing angle in a completely different direction. This discrepancy in viewing geometry means that the apparent position of features is physically shifted, causing pixel locations to be displaced on the Earth’s surface [26,27]. Figure 1 illustrates this effect, showing how a single fire event observed by both LEO and GEO satellites from different viewing angles results in an apparent spatial displacement of the fire location. Beyond the parallax effect, another major source of spatial misalignments is the sensor’s Point Spread Function (PSF) and the pixel remapping process [5]. The Point Spread Function (PSF) is an inherent optical property of the instrument. Due to diffraction, the radiant energy of a fire disperses outside the center field of view, spreading the thermal signal into adjacent detectors. In addition, the remapping operation of satellite pixels onto a fixed Earth grid requires interpolation, which smears the radiant energy of a sub-pixel fire across adjacent pixels, further distorting its apparent location.

Figure 2 shows an example of misalignments between the locations of fires detected from GEO and LEO images in three regions—California-Oregon, Amazonas, and Patagonia. The GEO images were captured by the GOES 16/17 ABI, whereas the VIIRS instrument acquired the corresponding LEO images. The figure shows that, although both instruments observe the same fires, there are significant spatial displacements of several kilometers in their locations. Hence, blindly using LEO detections as ground truth may be highly problematic for evaluating GEO fire detection algorithms.

An additional complication is that misalignments between the locations of GEO and LEO fire detections do not follow a systematic pattern. This is illustrated in Figure 3, which shows, for two different fire events in the Amazonas, their locations as detected by GOES 16 ABI and VIIRS. Even though the two fires, which occurred on 17 August 2022, were approximately in the same region, their ABI-VIIRS displacements are in different directions (east direction in one, and north direction in the other). This is just one example, as similar inconsistencies appear in all regions considered in this study.

These misalignments raise important questions at the focus of this study: How common are misalignments between ABI and VIIRS fire detections? And how should the accuracy of a fire detection algorithm be assessed in light of such misalignments?

Several studies that evaluated fire detection methods using LEO detections as ground truth did not account for these misalignments [11,13,28]. In these studies, a GEO fire detection is considered a false alarm if there are no LEO detections inside the GEO pixel. Other studies applied either a circular buffer around each GEO fire-detected pixel [18,21,22,23] or a square window [20,24,25]. In these studies, a GEO fire pixel is considered a correct detection if there is at least one LEO fire detection in its buffer area. Otherwise, the GEO fire-detected pixel is labeled as a false alarm.

The primary objective of this study is to quantify the prevalence of GEO–LEO misalignments and assess their impact on the evaluation of GEO fire detection accuracy. To this end, this study uses the GOES-16/17 ABI Fire Detection and Characterization (FDC) product as the GEO-based data source, and the VIIRS active fire product as the LEO-based reference in three different regions. This study shows that spatial misalignments between the location of fire detections from GEO and LEO sensors significantly impact the accuracy estimates of GEO-based fire detection algorithms. It is further shown that such misalignments commonly occur across various latitudes and ecosystem types.

2. Materials and Methods

This section outlines the data and the methods used to quantify the misalignments between GOES ABI and VIIRS fire detections and to evaluate the GOES-FDC algorithm. First, the study areas and data sources are described. Next, the pre-processing steps performed to spatially align these datasets are detailed, including corrections for the bow-tie effect, the rasterization of VIIRS fire detections, and the non-maximal suppression step. Subsequently, the methodology for quantifying spatial misalignments and evaluating algorithm performance using various buffer sizes is presented. All analyses in this study are performed at the pixel level.

2.1. Study Area

Three distinct regions of varying latitudes were included in the dataset of this study: high latitude north in the California-Oregon region (western United States), equatorial latitude in Brazil’s Amazonas State, and high latitude south in Patagonia (Figure 4). These areas all experience wildfires and have different environmental and climatic conditions.

2.2. Data Sources

For each of the three regions, a dataset for the whole year 2022 of fire product pairs from the LEO VIIRS and the GEO GOES-16/17 ABI was constructed (GOES-17 for California-Oregon and GOES-16 for the other two regions) to have temporally matching fire detections. Only GOES ABI images taken at ±5 min intervals from the corresponding VIIRS images were considered. Figure 5 illustrates the distribution of time offsets between the paired GOES-ABI and VIIRS images across the three study regions. As shown, the median time offset is 1 min, and it is assumed that fire progression within this time-frame would not impact the misalignment and evaluation analyses.

For VIIRS, the collection 2, level-2, VNP14IMG active fire product with 375 m resolution [29] was downloaded. Only “high” and “nominal” confidence VIIRS fire detections were retained, while “low” confidence cases were omitted.

For GOES ABI, the FDC product was downloaded [5]. The GOES FDC has several mask codes for its detected fire pixels, as detailed in

Table 1. GOES FDC relevant fire mask codes used in this work.

Pixel Mask Code	Pixel Definition
10	Processed fire
11	Saturated fire
12	Cloud-contaminated fire
13	High probability of fire
14	Medium probability of fire
15	Low probability fire
30	Temporally filtered Processed fire
31	Temporally filtered Saturated fire
32	Temporally filtered Cloud-contaminated fire
33	Temporally filtered High probability fire
34	Temporally filtered Medium probability fire
35	Temporally filtered Low probability fire

. As the table shows, some fire pixel categories are referred to as “temporally filtered”. A fire-detected pixel is considered temporally filtered if it was also detected as a fire at least once in the past 12 h.

The categories of fire detection with the lowest false alarm rates are “processed” and “saturated”, both temporally and not temporally processed [5,18]. Thus, in this study, only pixels with the corresponding mask codes (10, 11, 30, and 31) were used to assess the commission error and the prevalence of misalignments.

The constructed dataset contains 156,952 VIIRS (

Table 2. Number of VIIRS fire detections in each region.

Region	VIIRS Fire Detections
California-Oregon	48,242
Amazonas	98,726
Patagonia	9984

) and 59,522 matching GOES FDC fire detections, whereby 14,645 of them are from the “processed”, “temporally filtered processed”, “saturated”, and “temporally filtered saturated“ categories (

Table 3. Number of GOES FDC fire detections divided by categories in each region.

Region	Processed	Saturated	Cloud Contamination	High Confidence	Medium Confidence	Low Confidence	Total
California-Oregon	2123	14	1262	484	70	1802	5755
Amazonas	12,280	22	3993	1109	221	35,456	53,081
Patagonia	206	0	77	73	16	314	686

). The dataset size is comparable to that used for the GEO fire detection evaluation [13,18]. Appendix A and Appendix B provide additional details about the fire products and their source materials.

2.3. Removal of Duplicate VIIRS Detections—The Bow-Tie Effect

The VIIRS instrument has a whiskbroom scanner, where pixel size increases toward the edges of each scan. Therefore, pixels near the edges overlap with those from previous scan lines, resulting in pixel overlap. This phenomenon is known as the ”bow-tie effect” [30]. VIIRS removes some of the duplicated pixels in its standard processing pipeline. However, several residual “bow-tie” pixels remain. These duplicates may lead to an incorrect estimation of the probability of fire detection by GOES FDC. Hence, these VIIRS duplicate fire detections need to be removed. This issue is detailed in Appendix C.

2.4. Adjustment of VIIRS Pixel Size as a Function of View Zenith Angle

The nominal spatial resolution of VIIRS active fire pixels is 375 m at nadir. However, the actual pixel size increases with the View Zenith Angle (VZA) of the VIIRS scan. To account for this increase, the actual pixel size in meters ($S_{act}$) was calculated as a function of the VZA [31] as follows:

$$S_{act}={\displaystyle \frac{S_0}{cos\left(\theta \right)}}$$

(1)

where $S_0$ is the nominal pixel size at nadir (375 m), and $\theta$ is the VZA in degrees. This correction ensures that the spatial resolution of each VIIRS fire detection reflects its true surface coverage. The corrected pixel sizes were subsequently used during the rasterization of VIIRS detections onto the GOES-ABI grid.

2.5. Rasterization of VIIRS

As VIIRS and GOES ABI have different spatial resolutions and projections, the VIIRS data were rasterized into the GOES ABI grid. Specifically, the size-adjusted VIIRS fire detections were rasterized into the GOES ABI grid using the “rasterize” function from the “Rasterio” Python library (version 3.9.23) [32]. The rasterized output is a GOES ABI grid where each pixel is assigned the number of VIIRS pixels that intersect it. A GOES ABI pixel was assigned a zero value if it had no intersection with VIIRS pixels. Figure 6 illustrates this process.

2.6. Prevalence of Misalignments

The prevalence of misalignments between the locations of fires detected by GOES FDC vs. VIIRS in a given dataset is estimated as follows: Let m be the total number of FDC fire detections in the dataset. Each detection $1\le k\le m$, located at GOES ABI grid point $(i,j)$, is assigned to one of three possible categories: If there is at least one VIIRS fire detection inside the FDC pixel $(i,j)$, then the pixel is assigned to the first category with $GF{P}_k^{in}=1$. Otherwise, $GF{P}_k^{in}=0$, which may be due to misalignment. In this case, the presence of VIIRS fire detections is checked in a $3\times 3$ buffer window around the pixel $(i,j)$. If there is a VIIRS fire detection inside this buffer, then $GF{P}_k^{out}=1$ (category two); otherwise, $GF{P}_k^{out}=0$ (category three). Subsequently, the following sums are computed:

$$V_{in}=\sum _{(k=1)}^{m}GF{P}_k^{in}$$

(2)

$$V_{out}=\sum _{(k=1)}^{m}GF{P}_k^{out}$$

(3)

Next, the total number of GOES FDC false alarms is computed:

$$G_{FA}=m-(V_{out}+V_{in})$$

(4)

Finally, the percentages of these three categories are computed as follows:

$$\mathrm{\%FDC\; fire\; detections\; with\; VIIRS\; detections\; inside}={\displaystyle \frac{V_{in}}{m}}\times 100$$

(5)

$$\mathrm{\%FDC\; fire\; detections\; with\; VIIRS\; detections\; around}={\displaystyle \frac{V_{out}}{m}}\times 100$$

(6)

$$\mathrm{\%FDC\; false\; alarm\; rate}={\displaystyle \frac{G_{FA}}{m}}\times 100$$

(7)

Equation (6) provides the percentage of FDC fire detections that are misaligned with those of VIIRS.

2.7. Misalignment-Aware GOES FDC Fire Evaluation

This section describes a misalignment-aware methodology for evaluating the accuracy of a GEO fire detection algorithm using LEO-based detections as ground truth. While this study applies the evaluation to the GOES-16/17 FDC product and VIIRS, the methodology is flexible and can be applied to other GEO/VIIRS combinations. The corresponding Python code is publicly available at the following GitHub repository (https://github.com/asafvanunu/GEO_LEO_evaluation, accessed on 1 March 2026). Specifically, as described in Section 2.2, the assumed input is a dataset of GEO/VIIRS images of the same region, observed at matching times (within a ±5-min difference).

The first step is to use the VIIRS detections to construct a ground-truth reference of FDC fire pixel locations for each FDC image in the dataset. As discussed in several works [13,18,22], due to the higher spatial resolution of VIIRS, it can detect small-sized fires much more accurately than the GOES FDC. Hence, when evaluating the GOES FDC product, one user-defined parameter is the threshold t. This threshold specifies the minimum number of VIIRS fire detections required within an FDC pixel for it to be labeled as a fire pixel. Indeed, one quantity of interest in evaluating the GOES FDC output is its probability of detecting a fire, depending on its area, as measured by the number of VIIRS detections of the fire, for example Li et al. [13].

The fact that there may be spatial misalignments between the fire locations of the FDC product and VIIRS makes this evaluation not straightforward. Suppose, for example, that we wish to estimate this probability with at least $t=5$ VIIRS detections. Due to spatial misalignments, the VIIRS detections may be spread over two adjacent FDC pixels. This is illustrated in the bottom right corner of Figure 7 (left panel), where 6 VIIRS detections are spread with 4 in one FDC pixel and 2 in an adjacent one. To account for this issue, the following non-maximal suppression process is applied to the output image of the rasterize function in Section 2.5. We note that non-maximal suppression is a common set of techniques widely used in low-level computer vision tasks, particularly to handle sub-pixel misalignments [33]

In our setting, let $v(i,j)$ be the number of VIIRS fires inside an FDC pixel in a given location, also denoted as a cell. Let t be the user-chosen threshold for the number of VIIRS detections. A cell with a value $v(i,j)\ge t$ is labeled as a fire pixel. To address sub-pixel misalignments, if a cell has a value $0<v(i,j)<t$ and a neighboring cell $(i^{\prime },j^{\prime })$ has fewer but non-zero number of detections $0<v(i^{\prime },j^{\prime })<v(i,j)$, then the value at the original cell is updated to their sum $s=v(i,j)+v(i^{\prime },j^{\prime })$. If $s\ge t$, the cell is labeled as a fire pixel. If two neighboring cells have the same number of VIIRS detections, $v(i,j)=v(i^{\prime },j^{\prime })$, and $v(i,j)+v(i^{\prime },j^{\prime })\ge t$, they are both labeled as fire pixels. Such duplicated fire pixels are treated in post-processing to ensure no double-counting when calculating the omission error rate. The precise details of this non-maximal suppression step are described in Appendix D. The rasterized image of VIIRS fire detections after non-maximal suppression is considered as the reference ground truth. Figure 7 illustrates the process for a VIIRS rasterized image as input and a threshold $t=5$. With the non-maximal suppression procedure, the resulting output image contains eight fire pixels, whereas only three are present without a correction.

Next, true positive (TP) and false positive (FP) FDC pixels are identified, taking into account the potential FDC-VIIRS spatial misalignments.

To interpret the spatial extent of these buffer windows, it is necessary to consider the varying pixel size of the GOES ABI sensor. The nominal 2 km nadir resolution degrades as the view zenith angle increases [34]. Consequently, the area covered by a specific buffer size varies significantly between the equatorial Amazonas region (∼2 km pixel size) and the higher-latitude regions of California-Oregon and Patagonia (∼4 km pixel size).

Table 4. Approximate ground dimensions (in km) of the buffer windows for each study region, based on the degradation of GOES ABI pixel resolution with view angle [34].

Buffer Size (Pixels)	Amazonas (∼2 km Pixel)	California-Oregon/Patagonia (∼4 km Pixel)
$3\times 3$	$6\times 6$ km	$12\times 12$ km
$5\times 5$	$10\times 10$ km	$20\times 20$ km
$7\times 7$	$14\times 14$ km	$28\times 28$ km

details the approximate ground dimensions for each buffer window size used in this study.

Accordingly, a buffer window is applied around each FDC fire detection as follows. Specifically, let $I^G$ be the FDC image, where $I^G(i,j)=1$ at pixel locations $(i,j)$ where the FDC product detects a fire and $I^G(i,j)=0$ otherwise. In addition, for a given threshold t, let $I^L$ be the VIIRS image after the rasterization and non-maximal suppression steps. An FDC fire-labeled pixel is considered a TP if the buffer around it contains a VIIRS fire-labeled pixel. The FDC pixel is regarded as an FP if no VIIRS fire-labeled pixels are present inside the buffer window.

Finally, false negative (FN) events are computed as follows. If the buffer window around a VIIRS fire pixel does not contain an FDC fire-labeled pixel, that FDC pixel is considered a FN pixel, also known as an omission error. For this computation, an FDC pixel is considered as fire if it belongs to any of the fire mask codes in Table 1. Because the non-maximal suppression step aggregates detections from adjacent grid cells, it can cause an artificial duplication of fire pixels. Specifically, if two adjacent cells contain an identical number of VIIRS detections and their combined sum exceeds the user-defined threshold t, both cells are updated to the aggregated value. If these adjacent cells are missed by the GOES FDC product, counting them separately would artificially inflate the omission error rate. To avoid this double-counting, a fractional FN weight of 0.5 is assigned to these aggregated cells. Consequently, the adjacent cells contribute exactly one FN to the final evaluation metrics. The precise details of this correction step are provided in Appendix D. The remaining pixels are considered true negative (TN).

The computation of FN, TP, and FP with a buffer window of size $3\times 3$ is demonstrated in Figure 8. The red pixels annotated as GFP are detections made by the FDC fire product. In contrast, the green pixels annotated as VFP are VIIRS potential fire pixels after the non-maximal suppression step.

2.8. Accuracy Assessment

The accuracy of a GEO fire detection algorithm is quantified by its recall, precision, and F1 score. These are common quality measures used to assess classifiers. In the context of fire detection, recall quantifies the ability of a GEO fire algorithm to detect the ground truth fire pixels. Precision measures the fraction of actual fire pixels out of all those declared as fire by the GEO algorithm. The F1 score is the harmonic mean of precision and recall. Each accuracy measure is in the range [0,1], with higher values indicating better performance.

Given the number of TP, FP, and FN pixels for a set of n pairs of FDC/VIIRS images, as computed in Section 2.7 above, the recall, precision, and F1 score are calculated as follows: First, the total number of TP, FN, and FP pixels across all images in the dataset is computed. Let k denote the index of a single image. The total values are

$$T{P}^{total}=\sum _{(k=1)}^{n}T{P}_k$$

(8)

$$F{N}^{total}=\sum _{(k=1)}^{n}F{N}_k$$

(9)

$$F{P}^{total}=\sum _{(k=1)}^{n}F{P}_k$$

(10)

The accuracy measures are then estimated by

$$Recall={\displaystyle \frac{T{P}^{total}}{T{P}^{total}+F{N}^{total}}}$$

(11)

$$Precision={\displaystyle \frac{T{P}^{total}}{T{P}^{total}+F{P}^{total}}}$$

(12)

$$F1=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

(13)

In addition, via a bootstrap procedure [35], 95% confidence intervals are computed for each measure with 1500 repetitions.

3. Results

3.1. Prevalence of Misalignments Between GOES FDC and VIIRS

Figure 9 illustrates the percentage breakdown of the three categories related to FDC fire detections, as discussed in Section 2.6. The ratios were computed separately for each of the three regions in our study. Remarkably, about 12% of the FDC fire detections in all three regions have misalignments with VIIRS fire detections. These misalignments were observed in both GOES-16 and GOES-17 FDC images, across various regions, latitudes, and ecosystem types. As will be shown in the next sections, this high rate of spatial misalignments has important implications for the evaluation of the accuracy of the GOES FDC.

3.2. GOES FDC Precision, Recall, and F1 Scores

As described in Section 2.7, an FDC pixel is labeled as containing fire if at least t VIIRS detections are inside it. To provide a comprehensive evaluation across different fire sizes, the accuracy of the GOES FDC was evaluated across four threshold values: $t=1,3,5,$ and 7. For each value of t, four buffer window sizes were considered: $1\times 1$ (no buffer), $3\times 3$, $5\times 5$, and $7\times 7$ pixels.

Figure 10 displays the precision, recall, and F1 scores of the GOES FDC as a function of the threshold t for the three study regions. A general trend across all regions is that as the threshold t increases, precision declines. This occurs because when t is higher, GOES ABI pixels with fewer than t VIIRS detections are not considered as fire, and thus, valid GOES detections of small fires are counted as FPs. Conversely, as t gets higher, the recall and F1 scores generally increase. This demonstrates that the GOES FDC struggles to detect small fires. However, the FDC detection performance improves significantly for larger fire events. While the general trends hold, there are notable exceptions at the highest thresholds. In Patagonia, between $t=5$ and $t=7$, there is a decline in precision with the $1\times 1$ window. With the $3\times 3$, $5\times 5$, and $7\times 7$ windows, the precision remains relatively stable. This might be because there are not many fires in that region with more than seven VIIRS fire detections inside an ABI pixel. In the Amazonas region, the F1 score for the $1\times 1$ window declines between $t=5$ and $t=7$, whereas for the $3\times 3$, $5\times 5$, and $7\times 7$ windows, the F1 scores remain almost the same. This plateau and decline in the F1 metric are most likely driven by the sharp drop in precision values between $t=5$ and $t=7$, which occurs because there are many VIIRS fire detections inside GOES pixels with a count of less than seven.

The results also illustrate the impact of spatial misalignments. For instance, for $t=1$, without a buffer window ($1\times 1$), the estimated precision ranges from approximately 0.74 to 0.79. Since the false alarm rate equals one minus the precision, this corresponds to a high estimated GOES FDC false alarm rate of roughly 21% to 26%. However, this estimate is highly misleading due to the non-negligible percentage of misalignments. Taking a $3\times 3$ window yields a precision of around 0.85 to 0.95, significantly reducing the apparent false alarms. This follows the percentages shown in pink in Figure 9.

Regarding recall, the two regions of California-Oregon and Patagonia behave similarly, whereas the Amazonas has a significantly higher recall value. The reason for this, as shown in Table 3, is that in the Amazonas region, there is a much higher number of GOES FDC pixels classified as low confidence. This, in turn, implies that the number of FN pixels in the Amazonas region is relatively low, which yields these higher recall values. In all panels, the Patagonia region exhibits wider confidence intervals due to the smaller number of GOES FDC fire detections in that region. Overall, these results illustrate the importance of accounting for spatial misalignments between FDC and VIIRS detections in evaluating the GOES FDC product.

3.3. GOES Fire Detection Probability

Next, the probability of the GOES FDC to detect a fire as a function of the number of VIIRS fire detections in a window buffer around it was analyzed across the three study regions. Similarly to Section 2.7, let $I^G$ denote the binary image with GOES FDC fire detections. For a window width w, let $N_{ij}$ be the set of all neighboring pixels in the $w\times w$ window centered at $(i,j)$.

$$N_{ij}=\{(i^{\prime },j^{\prime })\mid max\left(\right|i^{\prime }-i|,|j^{\prime }-j\left|\right)\le w\}$$

(14)

For various values of an integer k, the total number of pixels $N_{TP,k}$ with at least k VIIRS detections and an FDC fire detection in a window buffer around them, an FDC/VIIRS image pair was computed across all FDC/VIIRS image pairs in the dataset. Specifically, for each image $I^G$, let $I^L$ be the output of the rasterize step with a parameter k, for the corresponding VIIRS image. Then,

$$N_{TP,k}=\sum _{images}\sum _{(i,j)}1\left[I^L(i,j)\ge k \mathit{and} \mathit{for} \mathit{some}(i^{\prime },j^{\prime })\in N_{ij},I^G(i^{\prime },j^{\prime })=1\right]$$

(15)

Similarly, the number of false negatives was also computed,

$$N_{FN,k}=\sum _{images}\sum _{(i,j)}1\left[I^L(i,j)\ge k \mathit{and} \mathit{for} \mathit{all}(i^{\prime },j^{\prime })\in N_{ij},I^G(i^{\prime },j^{\prime })=0\right]$$

(16)

By the above definitions, $N_{TP,k}+N_{FN,k}$ is the total number of pixels that contain at least k VIIRS fire locations. The probability that in a $w\times w$ window around each such pixel, there is an FDC fire detection was estimated:

$$P(I^G(i^{\prime },j^{\prime })=1 \mathit{for} \mathit{at} \mathit{least} \mathit{one}(i^{\prime },j^{\prime })\in N_{ij}\mid I^L(i,j)\ge k)={\displaystyle \frac{N_{TP,k}}{N_{TP,k}+N_{FN,k}}}$$

(17)

Figure 11 depicts the resulting estimated probabilities as a function of k for three different buffer sizes. In the Amazonas region (Figure 11b), the detection probability increases with the value of k until it starts to decline at $k\ge 15$. This slight decrease is likely due to a diminishing number of observations at higher k values, which may lead to inaccurate estimates of the true detection performance. In California-Oregon (Figure 11a), the GOES FDC fire detection probability increases with the number of VIIRS detections k. In Patagonia (Figure 11c), similar to the California-Oregon and Amazonas regions, the probability of a fire detection also increases with k. Furthermore, the detection probability in Patagonia is substantially higher when utilizing a $3\times 3$ or $5\times 5$ window. Notably, between $15\le k\le 20$, the probabilities for the $3\times 3$ and $5\times 5$ windows are the same.

Importantly, in all regions, the probability increases with the introduction of a buffer window, showcasing the prevalence of misalignments between VIIRS and GOES detections. For instance, with $k=1$, the detection probabilities without a window ($1\times 1$) are only 0.15, 0.10, and 0.13 for the Amazonas, California-Oregon, and Patagonia regions, respectively. In contrast, when applying a $5\times 5$ window, these probabilities increase to 0.36, 0.25, and 0.28, respectively. In addition, as k increases, the number of pixels with at least k VIIRS detections decreases sharply.

4. Discussion

In this study, VIIRS fire detections were used to evaluate the performance of the GOES-16/17 ABI FDC product. However, such an evaluation is not straightforward due to spatial misalignments between GEO and LEO fire detections. To address this issue, this study presents a scheme that incorporates a buffer window and a non-maximal suppression step to account for these misalignments. The scheme is generalizable and can be applied to other GEO/VIIRS sensor combinations, as most GEO systems share similar spatial, multispectral, and temporal characteristics. This study’s results demonstrate the importance of this approach in mitigating the impact of spatial misalignments.

To evaluate the GOES FDC performance, a large dataset of fire detections was constructed, covering three regions of different latitudes and containing approximately 157,000 VIIRS fire detections and 15,000 GOES FDC fire pixels (including processed and saturated categories). The evaluation process treats each pixel-level detection as an independent observation. While sequential detections may correspond to the same fire event across time, each acquisition is treated independently as it represents a new image taken by the satellite.

A key finding is that FDC-VIIRS misalignments occur in approximately 12% of the GOES FDC fire detections. This holds for both GOES satellites (GOES 16/17) and across latitudes and ecosystem types. To the best of our knowledge, no previous work has quantified this issue.

A second key finding is the importance of incorporating a buffer window around the FDC fire detections. For example, without a buffer, the precision of the GOES FDC in the three study regions is estimated to be in the range of 0.74 to 0.79. In contrast, with a buffer, the estimates are significantly higher, ranging from 0.85 to 0.93. In addition, implementing a buffer window yields significantly improved recall and F1 scores estimates, effectively mitigating the impact of FDC-VIIRS misalignments. These findings underscore the need to account for misalignments to ensure reliable fire detection assessments.

The precision values remain relatively constant for buffer window sizes $3\times 3$ and higher. This finding indicates that pixels falsely detected as fire by GOES FDC (without matching VIIRS detections) were indeed FP. The $7\times 7$ buffer window achieves the highest precision, recall, and F1 score. However, applying such a large window increases the risk of matching unrelated fires. As shown in Table 4, in high-latitude regions like California-Oregon and Patagonia, a $7\times 7$ window covers a substantial area of approximately $28\times 28$ km. Matching fires over such large distances likely introduces false associations between independent fire events. Therefore, a smaller window size is recommended to avoid such errors. Recall values are low when using a VIIRS fire detection threshold $t=1$. This is regardless of the buffer window size. These low values highlight the limitation of the GOES FDC to detect small fires.

In light of the findings above, the use of a spatial buffer window to evaluate GOES FDC performance is highly important. As discussed in Section 1, direct GEO/LEO pixel-to-pixel comparisons are inherently complicated by geometric and optical distortions. Standard ABI data currently lack terrain correction, meaning the parallax effect can induce apparent geolocation offsets of several kilometers in different regions [26,27]. Furthermore, the sensor’s PSF smears a sub-pixel fire’s radiant energy across adjacent pixels. The GOES FDC algorithm explicitly assumes a PSF where only 75% of the signal originates from the center field of view for the 3.9 $\mathsf{\mu }$m band, and only 51% for the 11.2 $\mathsf{\mu }$m band [5]. This dispersion is further exacerbated by the data remapping process. Therefore, a spatial buffer is required to mitigate spatial uncertainty and ensure an accurate assessment of the true detection capability of the GOES FDC product.

The large percentage of spatial misalignments in FDC/VIIRS image pairs has significant implications for the evaluation and training of machine learning-based GEO fire detection algorithms. Typically, VIIRS data are used as ground truth for training GEO-based machine learning fire detection algorithms [10,11,36]. However, to the best of our knowledge, most machine learning works to date have not considered misalignments between GEO and LEO sensors. Constructing a ground truth reference without taking the GEO/LEO misalignments introduces label noise into the training dataset. Standard training on a dataset with label noise often yields sub-optimal classifiers [37]. Hence, there is a need to account for spatial misalignments also in training machine learning models. Another interesting future direction is to extend this beyond pixel-level evaluation. The methodology proposed in this study can be utilized for burned area analysis, using aggregated fire pixels as an estimation for the total burned area. This transition from pixel-level validation to event-level assessment would provide a more comprehensive understanding of fire dynamics over time.

5. Conclusions

This study establishes that spatial misalignments between GOES ABI and VIIRS fire detections are a systemic issue, occurring in 12% of cases regardless of latitude or land cover. The results demonstrate that a direct pixel-level evaluation is insufficient for assessing GEO fire products due to inherent geometric and optical distortions. The implementation of a spatial buffer window is necessary to mitigate the effect of misalignments and estimate the true detection capability of the GOES FDC product.