Evaluation of Multi-Spectral Band Efficacy for Mapping Wildland Fire Burn Severity from PlanetScope Imagery

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper presents an approach to mapping burned area and burn severity (mapping of black vs. white ash). Research seems sound, the paper is is well structured and well written, and I have only minor observations.

Abstract makes reference to hyperspectral drone imagery but it is not explained how this relates to the research at hand (drones used in previous research, and as a validation source here).

Abstract makes reference to biomass consumption, but not calculation or validation of biomass consumption is presented.

The abstract should be reformulated and extended a bit to better reflect the research and main findings.

Introduction introduces problem well. Mention that Landsat repeat cycle has reduced with launch of Landsat 9. Mention that In the realm of free and open data, sentinel 2 offers both higher resolution and higher temporal coverage than Landsat.

If burn severity is introduced as biomass consumption, proof is needed that it is actually related to biomass consumption in the area under investigation. The cited reference (a government report) does not provide evidence. Furthermore, it does not seem to be publicly available (the URL only points to an abstract). Since in the further text, no reference to biomass consumption is made, reformulate accordingly. Further background should be presented to the reader on burn severity terminology (also for readers outside of the US with correpsonding if possible peer-reviewed references ). Anyway, the paper doews not measure fuel consumption but white and black ash.

In the imagery section, add a table with acquisition dates of both the drone and the satellite imagery, so the temporal difference can be clearly seen (this is important later for the discussion section on white ash).

In the conclusions or discussions section, ad a few sentences on options for using coarser resolution open data, especially Sentinel 2 (the table on imagery makes reference to the Planet Scope compatibility to Sentinel).

Minor language errors:

The SVM has been used in the previously in this research effort to determine the l. 98

Should probably read:

The SVM has been used previously in this research effort to determine the

Error! Reference source not found. shows the com- 201

PlanetScope imagery was acquired for each of the fires Planet Lab’s Education and l. 160

Should be

PlanetScope imagery was acquired for each of the fires through Planet Lab’s Education and..

Author Response

See attached pdf

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

L15 – ‘In addition to having higher spectral resolution and extent,’ – change spectral to spatial as you are saying in addition to spatial resolution, there’s also increased spectral resolution

L27 – identify whether the classification is supervised or unsupervised

The first paragraph of the introduction should be moved to the last paragraph of the introduction. It’s better to start an introduction with the broader context of the study, leading to set up why and what you’re going to do in the study at the end of the introduction

L38-40 It’s not as simple as ‘fire suppression driving the significant increase in large catastrophic fires’ in western US and the citation of grey literature is not good enough. There is a plethora of recent peer-reviewed scientific literature on the drivers of extreme fire activity in western US. Reference to the predominant effects of wind-driven and climate change drivers should be made with appropriate citations eg  https://doi.org/10.1029/2021GL095496     https://doi.org/10.1029/2022EF003471 https://doi.org/10.1002/fee.2359   https://doi.org/10.1186/s42408-019-0041-0

L71 This just adds confusion to terminology. See Keeley 2009 for recommended standardisation in terminology. Fire severity is a metric of the loss or change in organic matter caused by fire, which differs to fire intensity which is the energy output of the active fire.  https://doi.org/10.1071/WF07049 The suggested usage for your application is actually ‘fire severity’. Burn severity may be confused or conflated with what’s generally considered to be broader, longer term ecosystem responses. I suggest you should align with the suggested usage in Keeley 2009 and citing appropriately.

L77-82 Ash is generally not as reliable as looking at change in vegetation chlorophyll in remote sensing of fire, as ash may be contaminated or soon washed away before the postfire image is captured. The citation used is field oriented and not that relevant anyway. Acknowledge the large body of literature identifying the differenced normalised burn ratio and the RdNBR as the gold standards in remote sensing of fire extent and severity.  The lack of SWIR band in planetscope does limit the application of NBR derived indices, but this needs to be acknowledged, as perhaps a defensible reason to pursue the ash dominated/end members approach which has much less support and evidence in the literature.

L98 – typo – ‘used in the previously in this research’ change to ‘used previously in this research’

L108 – it surprises me that PCA would be required to reduce dimensionality when already using a machine learning algorithm, which can inherently efficiently handle very high dimensional data. SVM is particularly known for working well in high-dimensional spaces. Did you test the method without this PCA abstraction step?

L134 – ‘higher in the tree (closer to the root), while features with less information gain appeared lower in the tree (closer to the leaves).’ The analogy of closer to the root or closer to the leaves is confusing, as higher in the tree to most people would mean closer to the leaves. A tree’s root is at the bottom of the tree. Are they the wrong way around? If not, I’d just remove the words within the brackets in both cases.

L170 (FWHM) (ref) – remove (ref)

Table 1 is a direct reproduction of the table in the reference 28. Cite the reference in the table caption. This table could be extended to include the bands in sentinel 2 that are not present/interoperable in planetscope – namely the SWIR bands. This is an important point to make, as most recent satellite based fire severity mapping methods extensively rely on the NBR which uses the ratio of NIR to SWIR bands. The lack of SWIR bands is a limitation of planetscope imagery for fire mapping applications that needs to be acknowledged somewhere in the introduction

 

L201 - . Error! Reference source not found. Remove/fix citation error

Section 2.3.1 – rather than only using the raw bands as inputs to the model, there is a plethora of common vegetation indices with evidence in the literature showing efficacy in maximising the observation of fire effects. Even without the SWIR bands, there are plenty of strong candidate indices eg that compare Red and NIR in various ways, that should be tested for their information gain relative to the raw spectral bands. Eg the soil-ad[1]justed vegetation index, NDVI, burned area index, and even Tasselled-cap brightness and greenness transformations. This would strengthen the paper considerably and place it in the proper context of the current literature on remote sensing of fire. So you’d just generate the indices as a prior step then add them as additional bands in the input stack that you feed through your whole next process of comparing the information gains.

Figure 6 and 8 needs scale bar

Figure 7 needs scale bar and legend

Figure 9 caption is not clear, the insert overlay is a bit confusing.. Also no scale bar and it needs a legend describing the polygon colours. A bounding box around the drone imagery, and label it in the legend, would make this figure more clear.

 

L318-325 This whole method of further partitioning within the unburnt and burnt areas is artificial and unnecessary. Did you test whether it was actually ‘difficult for a SVM to identify the hyperplane separating burned and unburned classes due to a wide range of pixel values’? That would surprise me. If that is the case, then the main problem would be that you’re not providing enough richness of input data for SVM to work out the differences between burnt and unburnt. By providing vegetation indices as inputs, as mentioned previously, and not reducing the dimensionality of the data, SVM will do very well at classifying fire extent and severity.

 

Section 2.6. on validation data. Best practice and standard in the literature on these types of machine learning classifiers is to independently assess accuracy by training the model on a subet (eg 70% of the data) then testing accuracy on the remaining 30% that was not used to build the model. The way the validation data is constructed as a separate set of hand digitised polygons using maps already classified by the algorithm is not an independent test of accuracy.

 

Results Tables 5  to 11 – should be re-done using independently assessed accuracy, as described above. To more clearly examine differences between fires, why not combine all the fires for a full model, then compare those results to each fire by leaving the target fire out of the training data, train the model on all other fires and predict into the target fire.

Table 12 – averaging all the confusion matrices is very different to running the algorithm on the full set of data (leaving out the testing subset) and validating the full model on an independent subset, which is the far more rigorous and recommended approach.

Author Response

Please see attached pdf

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Please find the attached file. 

Comments for author File: Comments.pdf

Comments on the Quality of English Language

 Minor editing of the English language is required.

Author Response

Please see attached pdf.  Your comments were sent in a pdf from which we were not able to extract text to a google doc.  Consequently, we placed our responses in comments within the pdf you sent us.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have done a satisfactory job of addressing the concerns that had been raised.