Comments on "Systematically tracking the hourly progression of large wildfires using GOES satellite observations" by Liu et al.

The authors present the "GOFER" algorithm and derived data sets (GOFER-Combined, GOFER-West, GOFER-East) of hourly fire perimeters, active fire lines, and fire spread rates for 28 large wildfires in California from 2019-2021. Results are partially evaluated using a combination of fire perimeters from VIIRS data and final fire perimeters from California’s Fire and Resource Assessment Program (FRAP).

Specific Comments

* Section 1: Introduction

Line 59: "create a dataset" - Please choose another term than data/dataset and use throughout to make clear what are shown are not longer measurements, but results of calculations -  perhaps some type of "product"?

* Section 2.3 Using GOES active fire detections to derive hourly perimeters

Section 2.3.1: This section aims to provide a description of the GOFER algorithm but runs into some difficulty. The authors extend an approach originally described on a GEE blog (Restif and Hoffman, 2020) which here is cursorily described as a "GOES-based image-to-vector method to map fire perimeters in GEE". The remainder of the algorithm overview consists of describing optimizations/adjustments made to the original GEE approach clearly with the assumption that readers are already familiar with that approach. This is a mistake, I think, and more detail about the original method should be provided. The manuscript is cluttered with poorly defined and/or confusion quantities without this additional background material. An example is the "fire detection 'confidence' values" How were the arbitrary values in Table B1 selected? Why are saturated fire pixels assigned lower confidence? Table B3, which is deferred to the software (not algorithm) description, is of little help here, e.g., the fire confidence conversion "converts the codes indicating the quality of active fire detections to numeric values", but the fire detection codes are already numeric values.

Figure 2: This figure (upper right panel) implies that the authors assume there is zero geolocation error in the respective GOES ABI scans. This is not so, and the authors should include a brief discussion of the impact that these errors can have on their results. I believe they will find that this error is not much smaller than the downscaled 1.7 km resolution they claim to achieve by intersecting GOES-16 and GOES-17 pixels. A useful starting point might be the EMOSS team at NOAA (https://www.ospo.noaa.gov/Operations/GOES/goes-inrstats.html).

Section 2.3.2 "we create a dictionary of input values": Presumably "dictionary" (as opposed to, e.g., list or table) refers to the Python data structure. Clarify please.

Section 2.3.3.3: Please clarify if the "10 largest fires in California in 2020" are among those fires in the final 28-fire GOFER data set. Here and elsewhere the manuscript is unclear about which reference fires were used for parameter optimization and which reference fires were used to evaluate the final GOFER database. Presumably the authors are not using the same reference data for optimization and evaluation, but this needs to be made clear in the text.

Figure 4: IoU first appears here but is not defined. I am sure Intersection of Union is intended, and in this case the authors should clarify if they are calculating the metric with the actual perimeters or rectangular bounding boxes as is often the case in ML literature.

* Section 2.4 Derived fire metrics

2.4.1 Active fire line

Line 247: "We output fline_c at hourly confidence thresholds c of 0.05, 0.1, 0.25, 0.5, 0.75, and 0.9" It is unclear if this means that six different fline_c estimates are provided in the GOFER data set for each time step. If so, which one should those using the GOFER data select? Slightly more detail (but not enough) is given in the first paragraph of Section 3.3, but this should be included with the definition.

Line 262: "Thus, fline_r is more useful for studying the trends and behaviors of historical fires." The GOFER dataset only covers fires from 2018-2020 and is therefore inherently historical, so it is not clear why fline_c is provided ("cfinelen" in Table 2).

2.4.2 Fire spread rate

Here fire spread rate is defined in two different ways, but the authors do not provide a rationale. Please explain and, if relevant, discuss the conditions when each would be more appropriate for use. Which is at least theoretically more like to the spread rate as it would be measured on the ground?

It may be misleading to use the term "rate of spread" for what is being calculated, though that may be what the authors believe they are calculating.  For example, the (moderate) apparent growth outward of a perimeter can come about by rapid sliding downwind of air along the fire flanks, each parcel striping a bit farther out into unburned fuel. The rate any parcel of fire is spreading is then much different that the expansion rate calculated hourly normal to the fire line.

2.6 Evaluation and validation

Line 298: "FEDS can take advantage of the higher spatial resolution of 375-m VIIRS detections to nearly pinpoint the exact fire locations."  The more one uses VIIRS data, the less one tends to use terms indicating precision such as "pinpoint". While an excellent tool, similar problems arise when pushing VIIRS data beyond its limitations.

Please explain how comparing at 12 hourly intervals to VIIRS data or final perimeters can be said to "validate" hourly perimeters.

* Section 3.1 Evaluating the accuracy of the GOFER fire progression perimeters

Line 360: "In extreme cases, such as the Windy, Tamarack, Red Salmon Complex, and McCash fires for GOFER-Combined, we see this inability to form an initial perimeter hundreds of hours after ignition (Figure 4b)." These extreme cases comprise 4/28 = 14% of the GOFER data set. It seems appropriate to associate some sort of quality state to each GOFER fire and/or time step that would alert users to this situation.

* Section 3.2 The fire diurnal cycle derived from GOFER

Figure 9: What do the shaded area time series in the background represent?

Line 407: "The lower reliability of GOES-East during the day-to-night transition period likely drives the temporal artifacts in the fire diurnal cycle in GOFER-Combined." This explanation comes off as slightly misdirected because the authors are knowingly using GOES-East near the limit of its coverage.

* Section 3.3 Assessing the GOFER active fire lines and fire spread rates

Lines 443-449: The agreement between FEDS and GOFER fire line (fline_c and fline_r values) is fairly poor (average correlation from 0.45 - 0.64 depending on confidence threshold). FEDS is presumably better, though I don't believe the FEDS fire lines have actually been validated, but regardless the authors should discuss the implications of this result. Are the GOFER fire line estimates good enough to improve fire spread and atmospheric transport models as the authors claim?

Lines 456-463: The spread comparison is limited to the three different versions of the GOFER. The agreement is good but does not really say anything about the true accuracy of the spread rates. Also, the two different methods (MAE vs AWE) used to calculate spread rate show large differences (up to 2.5x). Is one more correct?

* Section 3.4 Future work and applications

Line 465: "The GOFER dataset can used to address key scientific questions on fire behavior controls in California." This and some of the claims that follow seem far fetched given the small sample size of GOFER (28 fires mostly from 2019 and 2020) and its inherently low spatial resolution.

Line 475: "the GOFER dataset can be used to build temporal or spatio-temporal statistical and machine learning models to understand how variations in climate, suppression, and fuels drive fire spread rate and fire-wide growth in area." The climate reference is excessive given the short time period covered by GOFER.

Line 478: "GOFER perimeters can be used to validate existing 3D fire spread models" At what model resolution do the authors think this might be true? The authors should keep in mind not only the low spatial resolution of GOES pixels but the non-zero error in the navigation of those pixels.

Overall:

This methodology and claims of what the work has accomplished is unsettling.   It not clear what scientific use one could legitimately use the product for, what spatial and temporal "resolution" one could claim it has, or what error one can associate with any given perimeter.  For example, there is so much texture to each hourly perimeter that is presented, yet (having studied several of these events in detail), nothing physical in fire behavior could support the bulges that the results show. (The unevenness of the perimeters likely arises from the error being of the order of the pixel size, as discussed elsewhere.) In contrast, the nature of the convection along the fire line has the opposite effect - to maintain a smoother fire line.

Review of essd-2023-389: “Systematically tracking the hourly progression of large wildfires using GOES satellite observations”

This is a well-written paper presenting an approach to map the hourly progression of the growth of 28 large fires that burned in California during 2019–2021 using the GOES-West and GOES-east satellites. Although the GOPHER fire perimeters are generally less accurate than those detected with higher-resolution imagery, the advance here is the high hourly time resolution. Some of the implications and potential uses for the current dataset may be a bit over-sold, but the dataset will certainly increase in value presuming that it grows in the coming years, and additional value will come when lessons learned with this approach are someday applied to new satellite products with higher spatial resolution and/or other improvements on GOES.

Specific comments:

L22: I realize that Intersection-over-Union (IoU) variable is a common metric in work related to image detection, but as a climate and fire scientist I was not aware of this metric until reading this paper and I suspect that many of the intended readers of this paper are similarly ignorant. For the Abstract, if a concise definition is not feasible, I think the main point that the fire perimiters detected in this study agree well with those from FRAP can be made without use of the IoU variable, or the meaning of the numbers could be made more intuitive. Then, in the main text I suggest explaining the IoU and any other metrics used in this study that may not be intuitive to fire scientists who lack expertise with remote sensing and/or image detection.

Section 3.4 (Future work and applications): While I do see value in this work, some of the suggested applications seem unlikely. For example, with only 28 fires, low spatial resolution, and uncertainty in the specific locations of burning and fire-line position, it seems doubtful that this specific dataset will open the possibilities for new insights regarding questions about the vegetation characteristics that promote explosive or quiescent fire activity. I think some of the caveats to this section, and any impression the reader may have that GOPHER is being oversold in this section, could be addressed if the section was preceded by a section that is specifically dedicated to the inherent limitations of GOPHER.

Technical corrections:

L249: “such as”?