Machine Learning Estimation of Fire Arrival Time from Level-2 Active Fires Satellite Data

Round 1

Reviewer 1 Report

This is a very interesting study employing a novel approach (SV machine learning) to conduct supervised classification in order to accurately interpolate wildland perimeters and spreads of wildfires for the purpose of initializing coupled weather-fire-smoke dispersion forecast models. Taking into account both fire- and clear-ground pixels in addition to satellite-generated confidence level of fire-pixel detection is something that has not been done before to my knowledge, and the results show that this approach is rather promising. I think this paper makes a significant contribution to the emerging field of coupled weather-fire behavior forecast models, and should be published.

I only have couple minor comments/suggestions:

1. The authors do not specifically state that the SVM algorithm did not use weather teaching data as input, although this is understood from the context of the paper, I think. It would be valuable to the reader to state this explicitly in the paper, and also comment on the limits of the algorithm in terms of time to correctly interpolate fire perimeters without the use of weather information.

 

2. The authors applied this SVM-based initialization approach to fires in CA. However, there is no discussion about how easy or feasible would be the application of the new method to fires in other states and across ConUS in general. The coupled weather-fire model WRF-SFIRE that the authors target with this study is intended to work at any location in ConUS. So, a broad applicability of the initializing procedure is important.
3. The text has a few minor typos, and requires a careful review by a technical editor to make sure that all sentences flow correctly and are free of typos.

Author Response

 

1. The authors do not specifically state that the SVM algorithm did not use weather teaching data as input, although this is understood from the context of the paper, I think. It would be valuable to the reader to state this explicitly in the paper, and also comment on the limits of the algorithm in terms of time to correctly interpolate fire perimeters without the use of weather information.
   
   Thank you for the suggestion. The SVM method is only using L2 Active Fires satellite data for the moment. However, we also proposed to use any other information of fire or no-fire in a specific location and time to improve temporal and spatial resolutions. For instance, one could use infrared fire perimeters used in the current paper for validation to improve the SVM estimation. As noted by the reviewer, the weather is not currently used in our analysis. However, the next steps would include using SVM results to initialize and assimilate a coupled atmosphere-fire model (WRF-SFIRE) which will be using and modeling the weather surrounding the fire. So, in our project, the weather is meant to play a role when forecasting, not when prescribing the fire evolution.
   
2. The authors applied this SVM-based initialization approach to fires in CA. However, there is no discussion about how easy or feasible would be the application of the new method to fires in other states and across ConUS in general. The coupled weather-fire model WRF-SFIRE that the authors target with this study is intended to work at any location in ConUS. So, a broad applicability of the initializing procedure is important.
   
   On line 498 of the paper, we have stated the following “…deliver global coverage. Therefore, the presented method can be applied globally.”. Since the products that our algorithm uses are global, we anticipate that the method presented is applicable across the globe.
   
   It is worth noting that this paper is not necessarily about WRF-SFIRE, as WRF-SFIRE is simply an example of a model that could benefit from the SVM method presented here. As noted in the text, we suggest that our SVM method could be used to downscale daily fire emission data to finer temporal intervals as noted on lines (94-99), among other applications.
   
   
3. The text has a few minor typos, and requires a careful review by a technical editor to make sure that all sentences flow correctly and are free of typos.
   
   Thank you for all your comments and suggestions. We have reviewed the draft to ensure that all grammatical and technical issues have resolved.

Reviewer 2 Report

This paper presents a novel fire propagation and estimated arrival time using support vector machine.  The authors use clear ground detection pixels to reduce false alarms from incorrectly classified fire detection pixels and to provide an estimate that can distinguish between an area that is not burning and a contiguous fire that may be partially obscured. Furthermore, the above sources do not generally utilize the confidence level associated with each fire detection, provided by active fires products.  The authors tested their method on 10 largest California fires during the 2020 fire season and evaluated with airborne fire observations.

 

The authors can improve their paper in two ways: consideration of other data sources and advanced methods:

 

Data sources:

The authors focused their methods only on MODIS and VIIRS.  There have been several approaches in literature that uses high resolution and high temporal data such as PlanetScope.  Furthermore, there are Geostationary data sources with higher temporal resolution that may be suitable for estimation aspects.

 

Methods:

Have the authors investigated recent advances in machine learning (deep neural networks) specially the ones relevant to time series?  There are numerous publications on using such techniques for detection and spread of fire using satellite images.

Author Response

Thank you so much for your comments and suggestions. Listed below are the responses that address the reviewer’s comments and suggestions.

Data sources:
Unfortunately, we looked at using PlanetScope early in our research process and found that it was not useful for our application. Since PlanetScope only measures visible and near-infrared parts of the electromagnetic spectrum, smoke aerosols often obscure the fire, which is a major limitation for our applications. However, we are currently working on incorporating GOES data into our procedure as a next step.

Methods:
We have not explored using deep neural networks since it requires a huge amount of data in order to work properly because of its model complexity. In our case, we were interested in using fire and clear-ground detection pixels from validated algorithms used in the community, which particularly worked better on machine learning methods with less complexity and that require fewer data. Also, SVM was chosen because of its definition of finding the best separation between two labeled groups which was exactly aligned with the goal of the project. It is worth noting that while deep neural networks are used as a surrogate of wildland fire spread models (trained using inputs and outputs from models to replicate their behavior), this was not the primary focus of our paper since our goal was to initialize fire locations with coupled fire-atmosphere models, air quality models, and to downscale daily fire emission data to a higher temporal resolution. For all these applications, we concluded that using a simple machine learning algorithm to estimate the fire evolution from satellite data was the best fit.

Reviewer 3 Report

Dear authors,

This paper proposes a new weighted classification machine learning method to estimate progression using both fire and clear ground detection pixels, in conjunction with each detection confidence level.

Below my comments:

Line 9 Infrared (IR) write the meaning of the acronym. In the abstract.

Line 23 should include a reference in the increase of fire activity.

Replace lines 122-129 for the main contributions of the research.

Methodology and results are well presented.

SVM method identifies fires spatiotemporally, using the fire arrival time concepts although the exploration of more meaningful ways to tune hyperparameters should be determined in order to improve the results.

Also, in the discussion part, should add some comparisons with similar studies mentioned in the introduction.

Author Response

• Line 9 Infrared (IR) write the meaning of the acronym. In the abstract.
  Thank you so much for your comment. “infrared” has been defined in line 9 in the text.
  
  
• Line 23 should include a reference in the increase of fire activity.
  Thank you for your suggestion. We have added the following reference on line 23:
  Spracklen, D. V., L. J. Mickley, J. A. Logan, R. C. Hudman, R. Yevich, M. D. Flannigan, and A. L. Westerling (2009), Impacts of climate change from 2000 to 2050 on wildfire activity and carbonaceous aerosol concentrations in the western United States, J. Geophys. Res., 114, D20301, doi:10.1029/2008JD010966.

• Replace lines 122-129 for the main contributions of the research.
  Thank you for your comment. The authors have included more detail in each section about the contributions:
  “The paper is organized as follows. Section 2.1 contains an overview of the analyzed fire events. The satellite data are described in Section 2.2, and the airborne fire observations used to validate our method in Section 2.3. The support vector machine (SVM) statistical learning method is reviewed in Section 2.4 where the novel weighted classification machine learning method is described. The core of the paper is the application of the SVM methodology to estimate the fire arrival time from satellite detection pixels in Section 2.5. The results for each of our case studies are shown in Section 3, where L2 AF satellite data is analyzed and fire arrival time estimation is evaluated using airborne IR fire perimeters suggesting that this method can be used to monitor wildfires in near-real-time. Section 4 summarizes the results presented throughout the paper and provides recommendations for future work.”
  
  
• Also, in the discussion part, should add some comparisons with similar studies mentioned in the introduction.
  Thank you for your suggestion. We added more content to lines 506-512 where now includes more comparisons with similar studies mentioned in the introduction:
  “In lieu of a typical spatial interpolation, the presented SVM method identifies fires spatiotemporally, integrating fire and clear ground detection pixels using the fire arrival time concept. The fire arrival time is calculated as the minimal time separating areas burning and not burning by using a cubic splines interpolation at each location. The satellite observations, which are discrete in time, are used to train the machine learning method presented here, which provides a continuous description of the fire progression. The fire extent can then be estimated at any given time in contrast with the similar studies mentioned in the introduction, which are designed to provide a daily estimate of the fire extent given the satellite data. Moreover, this novel method was proven to deal with false alarms or outliers by utilizing the confidence level associated with each fire detection pixel. The method also handles better small-scale irregularities since it provides a smooth estimation of the fire evolution for its kernel mathematical properties. Finally, the proposed method provides good quality estimates between satellite overpasses since it uses a weighted learning process not only utilizing the fire detection pixels but also the clear ground ones.”
  
  I would also like to clarify that a direct comparison of the statics between the results of this paper and the ones from the methods mentioned in the introduction can be misleading since they are using different cases and different ways to compute the statistics. However, we have been internally comparing our results with the ones provided in the introduction and our metrics look better than the ones in the papers. But, as mentioned, it is difficult to compare results from different studies. We are very grateful for the comments provided by the reviewer since they have enlightened us to start a discussion of how to test these other methods for upcoming research.

Reviewer 4 Report

It is a well documented article that I recommend for potential publication. However, may I suggest the authors to carefully look into the references, where I have seen duplications, e.g. [55] & [56], [53] & [54], [49] & [50], and others. 

Author Response

Thank you so much for your comments. The references listed here are slightly different, While they appear similar because they are from the same agency, these are technically different data sources.