Near Real-Time Fire Detection and Monitoring in the MATOPIBA Region, Brazil

Round 1

Reviewer 1 Report

The manuscript deals with an important topic in a sensitive region.  Unlike some portions of the tropics, this region experiences fires which are needed to preserve the savanna vegetation as well as fires which are not authorized and which lead to encroachment of agriculture into natural areas.  However, the manuscript is difficult to follow as described through comments in the annotated file attached.  In the concluding section, it is not clear how the results of this fire modeling method would be applied to improve management of both types of fires in the region. 

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comment

This manuscript discussed near real-time fire detection and monitoring in MATOPIBA region. This topic is interesting in remote sensing files. I reviewed it and provided some special comments for authors.

1. Title and Abstract are suitable for this manuscript.
2. In the end of Introduction (Lines 70-98), authors might consider to use concrete expression to show study purposes, not mixed too many information here, that might let audiences confuse. Moreover, the special and novelty of this research should be emphasized here in order to let audiences understand the contributions of this research.
3. Lines 100-102, Data chapter is too simple, more descriptions for that is necessary! Therefore, I suggested that authors should have more descriptions in this chapter.
4. Please confirm some Figures in this manuscript. Figures 6, 7, 9 and 10 are Tables? In my viewpoint, using Tables to express these Figures might be more suitable!
5. This manuscript combine Discussions and Conclusion as a chapter. Usually, they are individual. From the results of this study, audiences might not clearly understand that this approach is suitable or not suitable for MATOPIBA region compare to other regions (studies). Therefore, adding more references to discuss helps audiences to understand it.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript entitled “Near real-time fire detection and monitoring in MATOPIBA region, Brazil” presents the results of an applied research aiming to develop a model for fire detection based on GOES-16 ABI data. The authors use multiple methods to evaluate the performance of the models, including assessment of its performance in different land cover classes, with different fire sizes and different options regarding the number of consecutive AF detections. The developed model has an overall accuracy reaching 80% but the performance of the model when a single detection is adopted as AF detection drops to almost 57% with a high percentage of false positives. The latter is the main disadvantage on any automatic fire detection method. To be honest a find little operational value on this research especially if one takes into account that the already established models for fire detection and especially VIRS have similar or higher performance and with a higher spatial resolution.  Given that the highest accuracy of detection in the presented model is achieved after 125 consecutive detections then the advantage of using high temporal resolution data is  lost. Nevertheless, I have to acknowledge that the authors have done a lot of work on testing the potential of those data to be used for fire detection, while the methods employed are scientifically sound. So I believe the research should be published after some major improvements.

The introduction presents clearly the problem statement and the background regarding the need for timely and spatially accurate fire detection. I have to say that I strongly disagree with the authors regarding the dual role of fire in the area. Although any disturbance, including fire, tends to lead in increased biodiversity especially during the first stages of post disturbance secondary succession, in areas like the one presented in this study the main target should be to conserve the ecological integrity of natural and semi-natural ecosystems rather than the short-term increase of biodiversity. Furthermore, the continuous use of fire for maintaining the productivity of rangelands eventually leads to significant degradation due to soil erosion and often desertification, especially when it is combined with grazing.  We have witnessed this patter in Mediterranean region and elsewhere.  I would understand the use of fire for maintaining biodiversity in the Calcareous grasslands of UK, but for the tropical and sub-tropical ecosystems, I believe fire is a significant degradation factor. However, it is not me who is writing the manuscript and as a scientist I have to respect other scientist’s views even if I disagree. However, I had to make this comment just in case the authors wish to reconsider. Finally, I would either move the last paragraph of the introduction before the aim and objectives of the study or as the first paragraph of the methods section.

The methods are described in details and although the writing could be improved, they are clear and transferable.

The results sections needs significant improvement because it repeats a large number of numbers in the text which are already presented in tables.  All figures need to be improved in quality while most of them need to be converted to tables.

The discussion is rather general and not targeted on the results presented. I believe the authors should work on this section and focus more on the previously reported results.  

I have made some suggestions throughout the manuscript which the authors may wish to consider when revising their manuscript.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

This study aims to develop, the first Machine 4 Learning (ML) algorithm based on Advanced Baseline Imager (ABI) onboard the new generation of Geostationary Operational Environmental Satellite-R (GOES-R), able to detect and monitor Active Fires (AF) 5 in NRT in MATOPIBA. The paper analyzed the performance of 3 ML algorithms, and how many days are indicated to consider in a historical time series able to support accurate AF predictions. The most accurate algorithm was selected for the FM development. In this process, MODIS and VIIRS AF products were used for comparison as reference satellites and filtering purposes, and manually mapped BA on Sentinel-2 imagery.

The objectives of this work are to support the monitoring of fires in the daily activities of fire managers, comprehend FM potential and variables that influence its performance and characterize MATOPIBA’s fire based on the FM.

The introduction section of the manuscript is well written and documented, providing the background of the study. This section offers a good analysis about the fire incidence in different Brazilian biomes, especially in the tropical savanna Cerrado, which presents a singular dual relationship with fire, its incidence being necessary for biodiversity preservation, but when misused, has the potential to impact the environment. Nearly half of its original vegetation has already disappeared, mainly due to the advancing agricultural frontier in a region known as MATOPIBA.

The Data section is presenting the dataset used in this paper, composed by active fire (AF) products from the reference satellites (MODIS and VIIRS), GOES-16 ABI (Band 7) and the Sentinel-2 (Bands 4, 8A and 12) imagery, and Mapbiomas LULC mapping.

The Methods section has three sections, regarding data split, data processing and experiments, and FM development. In data split section, for the ML processes, the areas with burned area mapping were used, equally distributed among the three Natural Formation LULC, where 94 of the brightness temperature pixels were selected for the training, while 40 pixels were used for the test set and the remaining data pixels were used for the final inference process. For the data processing workflow the data normalization known as standard score (z-score) was applied, that requires not only the last brightness temperature, but also a historical time series. For the experiments, 10% of the training set was used, and created a z-score based on historical time series of 1 to 15 lags.

Integrated with the lags analysis, experiments were conducted with three different ML algorithms, aiming to identify the most suitable combination for the FM development: Random Forest (RF), Logistic Regression (LR) and Extreme Gradient Boosting (XGBoost). For each lag was applied each ML algorithm. Although in RF models were more efficient than LR for forest fire probability mapping, XGBoost presented an even higher performance. While RF and LR accuracy hardly achieves an accuracy of 60-70%, XGBoost achieves an accuracy of 70-80%. Due to the temporal resolution difference of the GOES ABI dataset (10 minutes), reference satellites (12 hours) and the BA mapping (5 days), FM accuracy in the test set was assessed considering both a single indication of AF and a certain sequence of AF. The consecutive AF indications means a more persistent fire along the time.

The results are presented divided by subheadings to provide a concise and precise description of the experimental results, their interpretation, as well as the experimental conclusions that can be drawn.

The Overall performance of the FM applied to the test set resulted in an accuracy rate of 78.9%. The probability of a FM being right when it points an AF detection is around 87%, and when it indicates a non-fire, around 70%. FM performance regarding Burned Areas Mapping, was found that the size of the BA does not influence the FM accuracy.

The accuracy rate of a single detection is 56.6%, 15 consecutives AF detections (after 2,5 hours) is 67.3%, and 125 (after ≈20 hours), achieves an accuracy peak of 73.4%. The accuracy rate of the reference satellites is almost 71% and roughly half of the fires are correctly detected. Besides, the reference satellites rarely commit false positives, less than 3%, yet its true positives are lower than the FM approaches. According to the BA manual mapping, almost 50% of the 125 consecutives AF detections from the FM and the reference satellites are correct and in agreement, and almost 5% are by both incorrectly classified.

The fire reality in the remaining data along MATOPIBA’s territory, comparing the fire prediction based on the reference satellites and the 125 consecutives AF detections, the difference between the results was only 167 AF in more than 2,400 pixels that were analyzed.

The discussion and conclusions section shows that the FM proved to be versatile among the three analyzed Natural Formations: Natural Forest, Savanna Formation, and Grasslands. The FM single detection accuracy can be negatively influenced by BA greater than 1 km² in the central pixels and in its surroundings, the greatest potential being when the fire is in its initial phase. FM consecutive AF was developed, and it presented a higher accuracy, reaching its peak after around 20 hours (125 consecutive AF). Even though there is an important tradeoff between the consecutive AF and time, the 125 consecutive AF presented a number of true positives almost as accurate as the reference satellites. FM was developed based on true or false prediction by means of the XGBoost model and also considering the z-score of the last 13 days. But, due to the seasonality throughout the year, it is indicated the retraining of the FM along the time. FM could be also improved by using a fire prediction confidence rate instead of the binary prediction (true or false).

The study is well written and documented and is an improvement of previous works. The methodology is not new, but authors show a good knowledge of using and improving it. The results obtained demonstrated the potential of the FM in the monitoring of fires.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have done a good job of responding to reviewer comments, providing a clearer manuscript that describes the approaches and findings in a logical way and linking to applications of the new methods.

Reviewer 2 Report

Comment for remote sensing-1668700R1

I reviewed the revised manuscript and found most of my suggestions has been improved. I appreciate that authors attentively reply each point and used to revise manuscript. I only have a slight suggestion for this version of manuscript. Usually, below figure should have some description, even though a short sentence. Such as in Figure 1, author might add some description for this Figure, such as “The detailed steps of Figure 1 were shown in below sections.”

Nevertheless, this small problem is easy to solve and I am pleased to recommend this manuscript for publication in remote sensing.

Comments for author File: Comments.pdf

Reviewer 3 Report

In my first review of the manuscript entitled in its revised form “Near real-time fire detection and monitoring in the MATOPIBA region, Brazil” I expressed some concerns regarding the operational value of the manuscript. However, I also acknowledged the scientific soundness of the study and its value to the scientific community.  

The authors in this revised version have done a good job regarding the content of the manuscript as well as its presentation format and structure. I am glad to say that they have addressed all my comments. Although as I expressed I still disagree with some views but a manuscript should express the views of the authors and not these of the reviewer. The results are now much more clearly presented and the discussion has been rewritten reflecting better the previously reported results.

With the above being said, I have no hesitation to recommend acceptance of the manuscript in its current form.