Object-Based Change Detection in the Cerrado Biome Using Landsat Time Series
Round 1
Reviewer 1 Report
Overall this is an interesting paper which examines an important issue of deforestation in the Cerrado which is very important and often overlooked. However, I do have three general concerns with it.
1. How saleable is your methodology and use of eCognition to larger areas?
2. OLI has 2 SWIR bands SWIR 1 [1.566 - 1.651 μm] and SWIR 2 [2.107 - 2.294 μm]. When addressing other reviewers’ comments, please clarify through the manuscript if you are referring to SWIR 1, SWIR 2 or bands in general with wavelengths between 1.5 to 2.3 μm.
3. I would like to see more discussion on how images were evaluated for suitability. I am guessing absence of cloud cover but this was not stated. Also, I would like to know if you gained any knowledge about the number and timing of images required for suitability of your time series methodology. If you wanted to do a detailed study on this, I think it could be a follow-up publication.
Abstract
In the abstract you say the following "methods that apply the spatial context have been used to improve the results of change detection. However, none of these studies has explored Landsat’s shortwave infrared channel (SWIR 2)." This statement is incorrect as many studies are published each month that use OLI SWIR 2 time series bands. Please rephrase.
Figure 1
Change R, G, B to Red, Green, Blue bands for clarity.
Figure 4
It is a bit difficult to distinguish Def and Sch in this figure.
Also, it is unclear what the darkest bars signify.
You include SWIR1 here along with SWIR2. Is this a typo?
Figure 6.
Was SWIR1 compared to the other bands? You mention it in Figure 4 but nowhere else.
Table 2.
I would really like to see accuracy assessments for each of the sites in addition to the overall accuracy assessment. Since you had different study areas which contained different input imagery, it would be useful to show the differences in accuracy between sites.
Discussions.
You mention limitations of study area size and number of images for time required for image processing. It would be helpful to know the approximate hardware specs and required processing time for each site. A detailed account of this information is not required, however including a small section on this in the results would be helpful to the reader.
Author Response
Response to Reviewer 1 Comments
English language and style: English language and style are fine/minor spell check required.
Response: The English language and style were revised by a native English speaker.
Overall this is an interesting paper which examines an important issue of deforestation in the Cerrado which is very important and often overlooked. However, I do have three general concerns with it.
1. How saleable is your methodology and use of eCognition to larger areas?
Response: Our method brings an option by using shortwave infrared channel (SWIR2) of Landsat 8 and object based image analysis to detect deforestation in Cerrado, a biome with high human-induced activities and seasonal variations. By using free satellite imagery and performing a simple change detection, this methodology can be explored by governmental agencies to provide a basis for systematic forest monitoring and conservation.
Regarding to larger areas, the object based change detection performed at eCognition may take an extensive computational time, but two points can face this issue. First, the multi-temporal segmentation is run by a single step, reducing for instance, time consuming by segmentation updates. Second, alternative segmentation methods, such as ENVI EX or even writing the multiresolution segmentation algorithm to another programming language, can be tested in order to improve the segmentation time performance in large areas.
2. OLI has 2 SWIR bands SWIR 1 [1.566 - 1.651 μm] and SWIR 2 [2.107 - 2.294 μm]. When addressing other reviewers’ comments, please clarify through the manuscript if you are referring to SWIR 1, SWIR 2 or bands in general with wavelengths between 1.5 to 2.3 μm.
Response: Agreed, addressed in revision.
Line 22: “… using the SWIR 2.”
Line 23: “…computed from SWIR 2...”
Line 25: “… The SWIR 2 channel …”
Line 124: “...Green = SWIR 1…”
Line 171: “…for the SWIR 2 spectral channel.”
Line 236: “…that the SWIR 2 band…”
Line 242: “…shown that SWIR 2 is a…”
Line 234: “…SWIR 2 using the…”
Line 263: “…derived from SWIR 2 temporal…”
3. I would like to see more discussion on how images were evaluated for suitability. I am guessing absence of cloud cover but this was not stated. Also, I would like to know if you gained any knowledge about the number and timing of images required for suitability of your time series methodology. If you wanted to do a detailed study on this, I think it could be a follow-up publication.
Response: Addressed in revision.
Line 109: “Thus, two prerequisites were determined for image selection: absence of clouds and presence of deforested areas and seasonal changes, which led to different deforestation rates and image frequencies for each sampled site.”
Line 281: “In other words, the image frequency can affect the number of deforested areas detected per site, which in turn, will affect the accuracy results. For instance, an omission error of one changed-object in a site where only two changed-objects were detected means an error of 50%, whereas in a site where five changed-objects were detected an omission error of one changed-object means an error of 20%.”
Abstract
In the abstract you say the following "methods that apply the spatial context have been used to improve the results of change detection. However, none of these studies has explored Landsat’s shortwave infrared channel (SWIR 2)." This statement is incorrect as many studies are published each month that use OLI SWIR 2 time series bands. Please rephrase.
Response: Rephrased.
Line 16: “However, few studies have explored Landsat’s shortwave infrared channel (SWIR 2) to discriminate between seasonal changes and land use/land cover changes (LULCC).”
Figure 1
Change R, G, B to Red, Green, Blue bands for clarity.
Response: Addressed in revision.
Line 122: “Figure 1. (a) The study area in the north of Minas Gerais (MG) State, Brazil, and the details of the sampling design; (b) cloud free image acquisition dates per sampled site; (c) sampled areas represented by Landsat OLI image false color composition (Red = NIR, Green = SWIR 1, Blue = Red) and Cerrado mask layer from land cover classification.”
Figure 4
It is a bit difficult to distinguish Def and Sch in this figure. Also, it is unclear what the darkest bars signify.
Response: Addressed in revision. We colored the bars in a distinguishable version, and added a label for overlapped bars between Def and Sch (darkest bars in the original figure).
You include SWIR1 here along with SWIR2. Is this a typo?
Response: No, it is not a typo. Since we evaluated OLI bands to determine the ability of each band to separate deforestation from seasonal changes, SWIR 1 and SWIR 2 must be evaluated separately as illustrated in Figure 4.
Figure 6.
Was SWIR1 compared to the other bands? You mention it in Figure 4 but nowhere else.
Response: The feature evaluation step was divided in two parts. We first compared OLI bands with each other to separate deforestation from seasonal changes (Figure 4). Second, as the SWIR 2 band presented the strongest Def/Sch separability, we select a set of spectral indices (based on SWIR 2 band) to detected changed objects (Figure 5) and computed the importance of each spectral variable (Figure 6). Thus, SWIR 1 was only mentioned in Figure 4 because further analyses were performed by SWIR 2 and derived indices.
Table 2.
I would really like to see accuracy assessments for each of the sites in addition to the overall accuracy assessment. Since you had different study areas which contained different input imagery, it would be useful to show the differences in accuracy between sites.
Response: Addressed in revision.
Line 194: “In addition, accuracy metrics were performed by each site separately.”
Site | Overall Accuracy (%) | Overall Error (%) | Producer’s Accuracy (%) | User’s Accuracy (%) | ||||||
Def | O.E. | Sch | O.E. | Def | C.E. | Sch | C.E. | |||
All | 88.3 | 11.7 | 88.0 | 12.0 | 88.6 | 11.4 | 84.6 | 15.4 | 91.2 | 8.8 |
1 | 83.3 | 11.7 | 75.0 | 25.0 | 100.0 | 0.0 | 100.0 | 0.0 | 67.7 | 32.3 |
2 | 87,5 | 12.5 | 100.0 | 0.0 | 75.0 | 25.0 | 57.1 | 42.9 | 100.0 | 0.0 |
3 | 93.3 | 6.7 | 100.0 | 0.0 | 87.5 | 12.5 | 87.6 | 12.4 | 100.0 | 0.0 |
4 | 86.7 | 13.3 | 80.0 | 20.0 | 90.0 | 10.0 | 80.0 | 20.0 | 90.0 | 10.0 |
Line 218: “The overall accuracy by site demonstrated low variability among all sites, with Site 3 presenting the highest value (93.3%) and Site 1 the lowest value (83.3%). The Def class presented good producer and user’s accuracies, except the user’s accuracy of 57.1% in site 2. Accuracies for the change maps of the sampled sites are shown in Figure 5.”
Line 276: “The overall accuracy analysis for each site demonstrated low variability (83.3 to 93.3%), while producer’ and user’s accuracies presented fluctuating measures for the Def class (75 to 100% and 57.1 to 100%, respectively). However, these accuracies per site are likely unreliable since the test set was randomly sampled from the four sites, thus the change-object frequency is most likely different for each site and may not be enough for a reliable site-specific accuracy analysis.”
Discussions.
You mention limitations of study area size and number of images for time required for image processing. It would be helpful to know the approximate hardware specs and required processing time for each site. A detailed account of this information is not required, however including a small section on this in the results would be helpful to the reader.
Response: Addressed in revision.
Line 310: “Even though the computational time in this study took less than a minute for each site to be processed by one computer, it is important to consider the sample size of 100 km² per site, which represents 0.2% of the Minas Gerais State and can take hours of processing the whole area.”
Author Response File: Author Response.docx
Reviewer 2 Report
A review of 'Object-based change detection in the Cerrado Biome using Landsat time series'
Summary
This study used multi-temporal object change analysis and Random Forest algorithm and a time-series of Landsat OLI data to detect deforestation in Brazilian savannas and distinguish it from phenological changes. The authors found SWIR and the Mid Infrared Burned Index (MIRBI) were the best variables for change detection and differentiating deforestation from seasonal changes.
The study is a welcome addition to the growing body of knowledge on combining spectral and spatial and phenological information for more effective LULCC mapping. Many change detection studies are based only on spectral information and only rarely incorporate object change analysis.
Overall Comments
The manuscript is well-written, and the analyses carried out are complex, and inevitably, the methods section will need a more systematic step-by-step description on how the inputs were prepared, and how the multi-temporal segmentation and feature extraction was carried out. Some of the steps described only reference Desclee et al. (2006) but the description that follows is confusing. The Methods section ill therefore needs elaboration and more clarity.
Specific Comments
Introduction
Line 68: 'motivated'
Line 69: 'proposed using the shortwave infrared'
Lines 74-76: Not clear what is meant. Could you rephrase and clarify this sentence.
Line 74: Replace 'Here' with 'In this study'
Material and Methods
Lines 106-108: 'cloud-free'. Also, why not just refer to 'Path' and 'Row' that are routinely used to identify locations of Landsat scenes? Why April 2013 to December 2014? What other considerations went into your choice of imagery besides 'cloud-free'?
Line 109: Could you clarify why top-of-atmosphere reflectances were not desirable? I am unfamiliar with 'bottom of atmosphere reflectance'. How different is it from 'surface reflectance'?
Lines 109-110: What impact does the number of images available for analysis at each sampling site affect your results?
Lines 111-112: Please provide some details on how the initial land cover map was generated and the acquisition dates of the imagery than went into it. How was the mask created, and more specifically, what LULC types were being excluded by the masks?
Figure (b): It would be important to come back to this figure in your discussions and explain what impact the frequency and seasonality of imagery analyzed at each sample site affected your results.
Lines 118-121: This is the section I had the most difficulty following. You need to include more details on how the segmentation was carried out. What imagery went into the multidate segmentation? How many 'bands' or variables were in each input 'layerstack' per sampling site? The label 'Image difference for all OLI bands' in figure (a) is rather confusing. What bands were differenced? Shouldn't the first step 'multidate segmentation' just mean 'partitioning of multidate image bands into groups of pixels that are spectrally similar and spatially adjacent'?
Figure (b): not clear how 'deforestation' and 'seasonal change' were distinguished.
Lines 124-125: What bands or image enhancements were differenced?
Lines 125-126: In the study by Desclee et al. (2006), multidate segmentation was used to partition the whole multiyear images into objects from which multidate signatures were extracted to characterize each object's spectro-temporal behavior. I did not see image differences as input images at this stage. Would you please clarify and elaborate?
Line 132: Are these the parameters found optimal after the 'trial and error'? What are some of the initial parameters tried and how did this affect the segmentation? It would be useful for the authors to describe their experience in experimenting with different parameters and the lessons learned so that other readers can tap into this experience into their own work.
Line 140: Figures (a) and (b). What bands are represented in these illustrations?
Results
Lines 226-228: Would the authors consider merging this one-sentence paragraph with others or expanding it? Perhaps they could explain why they think SWIR 2 band has the best potential to detect deforestation, yet it is not as good in distinguishing between deforestation and seasonal changes?
Lines 232-236: It is not clear what is meant by 'channels' being 'promisor'. Please rephrase this sentence and if possible, use another term in place of 'promisor'.
Line 258-260: Not clear why the tasseled cap coefficients used in study were inappropriate. Please elaborate.
Lines 272-273: What variation in number of images used is considered 'small', and what is an optimal number of images to include in a change detection?
Line 282: ' …...which have continuous data observations …..' Please rephrase sentence for clarity.
Lines 283-284: This is a 'one-sentence' paragraph that should either be combined with the previous paragraph or expanded.
Conclusions
Lines 287-289: But you stated earlier that the poorer ability to distinguish between deforestation and seasonal changes resulted in higher misclassifications between the two classes. Why the contradiction here?
Line 290: ' …. satisfactory change detection accuracies ….'
Author Response
Response to Reviewer 2 Comments
A review of 'Object-based change detection in the Cerrado Biome using Landsat time series'
Summary
This study used multi-temporal object change analysis and Random Forest algorithm and a time-series of Landsat OLI data to detect deforestation in Brazilian savannas and distinguish it from phenological changes. The authors found SWIR and the Mid Infrared Burned Index (MIRBI) were the best variables for change detection and differentiating deforestation from seasonal changes.
The study is a welcome addition to the growing body of knowledge on combining spectral and spatial and phenological information for more effective LULCC mapping. Many change detection studies are based only on spectral information and only rarely incorporate object change analysis.
Overall Comments
The manuscript is well-written, and the analyses carried out are complex, and inevitably, the methods section will need a more systematic step-by-step description on how the inputs were prepared, and how the multi-temporal segmentation and feature extraction was carried out. Some of the steps described only reference Desclee et al. (2006) but the description that follows is confusing. The Methods section ill therefore needs elaboration and more clarity.
English language and style: English language and style are fine/minor spell check required.
Response: The English language and style were revised by a native English speaker.
Specific Comments
Introduction
Line 68: 'motivated' Addressed in revision.
Line 69: 'proposed using the shortwave infrared' Addressed in revision.
Lines 73-75: Not clear what is meant. Could you rephrase and clarify this sentence.
Response: Addressed in revision.
Line 73: “Thus, change detection in the Cerrado biome is motivated by the previous assumptions, where an effective spectral analysis of the change matter is required to support change detection studies.”
Line 76: Replace 'Here' with 'In this study' Addressed in revision.
Material and Methods
Lines 100-102: 'cloud-free'. Addressed in revision.
Also, why not just refer to 'Path' and 'Row' that are routinely used to identify locations of Landsat scenes? Why April 2013 to December 2014? What other considerations went into your choice of imagery besides 'cloud-free'?
Response: The Landsat scenes locations was a typo and was addressed in revision.
We choose the specific two-year period based on four reasons:
(i) It started on April 2013 because it is the first month available for Landsat 8 image acquisition due to the satellite launch on February 2013. Also, this initial analysis period creates a starting point to further detections through Landsat 8 imagery.
(ii) With low computational effort, we can acquire more images per year and capture change events that may not be detected in biannual change method for instance, due to rapid land conversion in Cerrado regions;
(iii) In the multidate segmentation we avoid extensive processing time and the creation of small and spurious objects;
(iv) The method can generate an annual deforestation report providing support for a continuous forest monitoring and conservation program.
Beside cloud-free images, we chose images and locations with deforestation events to make sure we would have enough information to pursue further analysis.
Line 103: Could you clarify why top-of-atmosphere reflectances were not desirable? I am unfamiliar with 'bottom of atmosphere reflectance'. How different is it from 'surface reflectance'?
Response: Top-of-atmosphere (TOA) reflectance is measured by remote sensing instruments aboard satellites, and includes the reflectance from the surface and the reflectance from the clouds. The path of light through the atmosphere can change as the light travels down to the earth through the atmosphere, then diffusively reflects off of the Earth's surface, and back up through the atmosphere again, suffering wavelength-dependent scattering. Thus, TOA reflectance may represent a disturbed value from Earth’s surface and is not desirable in some applications. It can be transformed into Bottom-of-atmosphere (BOA) reflectance by several calculations, returning a pixel value without atmospheric noise.
There is no difference between the terms Bottom-of-Atmosphere Reflectance and Surface Reflectance.
Lines 103-104: What impact does the number of images available for analysis at each sampling site affect your results?
Response: The number of images available in a same period affects the results in two situations. Few images can omit change objects since the rapid land conversion and the Cerrado regeneration, i.e. if two consecutive images have a five months’ gap, a deforestation event can occur and the growth of natural regeneration can mask the changed area in a change detection analysis. On the other hand, many images can take extensive processing time and create small and spurious objects.
Line 103: “We set a short two-year period because: (i) we can acquire and process images with lower time gaps and lower computational effort compared to long time series; (ii) the method can detect change events that may not be captured by a biannual method due to a rapid land conversion, for instance; (iii) a short period can generate annual deforestation reports, providing support for continuous monitoring and conservation applications.”
Lines 117-118: Please provide some details on how the initial land cover map was generated and the acquisition dates of the imagery than went into it. How was the mask created, and more specifically, what LULC types were being excluded by the masks?
Response: Addressed in revision.
Line 116: “We generated a Cerrado mask based on a land cover classification using an object-based image analysis (OBIA) (Figure 1c). In this study, we applied the multi-resolution segmentation algorithm [41] implemented in the eCognition software [42], to create image-objects and classified them through Fuzzy logic and spectral parameter selections. Classification post processing, such as visual interpretation and manual editing, corrected misclassifications improving the final map accuracy.”
Figure (b): It would be important to come back to this figure in your discussions and explain what impact the frequency and seasonality of imagery analyzed at each sample site affected your results.
Response: Addressed in revision.
Line 276: “The overall accuracy analysis for each site demonstrated low variability (83.3 to 93.3%), while producer’ and user’s accuracies presented fluctuating measures for the Def class (75 to 100% and 57.1 to 100%, respectively). However, these accuracies per site are likely unreliable since the test set was randomly sampled from the four sites, thus the change-object frequency is most likely different for each site and may not be enough for a reliable site-specific accuracy analysis. In other words, the image frequency can affect the number of deforested areas detected per site, which in turn, will affect the accuracy results. For instance, an omission error of one changed-object in a site where only two changed-objects were detected means an error of 50%, whereas in a site where five changed-objects were detected an omission error of one changed-object means an error of 20%.”
Lines 127-128: This is the section I had the most difficulty following. You need to include more details on how the segmentation was carried out. What imagery went into the multidate segmentation?
Response: Addressed in the new Figure 2.
How many 'bands' or variables were in each input 'layerstack' per sampling site?
Response: Addressed in revision.
Line 132: “We created multi-temporal image-objects by segmenting multiple images difference of sequential periods. The procedure was conducted in a single operation using the whole set of spectral bands with sequential difference images together, which the layer stack dimension for each sampled site was determined by n-1 (site image frequency) multiplied by six (OLI bands) [43].”
The label 'Image difference for all OLI bands' in figure (a) is rather confusing.
Response: Addressed in the new Figure 2.
What bands were differenced?
Response: Addressed in the new Figure 2.
Shouldn't the first step 'multidate segmentation' just mean 'partitioning of multidate image bands into groups of pixels that are spectrally similar and spatially adjacent'?
Response: Yes. We adopted the term multidate segmentation as suggested by Desclée et al. (2006).
Figure (b): not clear how 'deforestation' and 'seasonal change' were distinguished.
Response: Addressed in revision.
Line 149: “Once the multidate segmentation was carried out, we labelled all image objects as Def or Sch by visual interpretation of time series.
Lines 132-133: What bands or image enhancements were differenced?
Response: Addressed in revision and in the new Figure 2.
Lines 132-133: In the study by Desclee et al. (2006), multidate segmentation was used to partition the whole multiyear images into objects from which multidate signatures were extracted to characterize each object's spectro-temporal behavior. I did not see image differences as input images at this stage. Would you please clarify and elaborate?
Response: Addressed in revision and in the new Figure 2
At this stage, image differences did compose the input image set. “The procedure was conducted in a single operation using the whole set of spectral bands with sequential difference images together”.
Line 143: Are these the parameters found optimal after the 'trial and error'? What are some of the initial parameters tried and how did this affect the segmentation? It would be useful for the authors to describe their experience in experimenting with different parameters and the lessons learned so that other readers can tap into this experience into their own work.
Response: Yes, these are the optimal parameters. A brief explanation about the tests performed were addressed in revision.
Line 138: “We applied the multi-resolution segmentation algorithm [41] using a trial-and-error approach [20,44] to find the appropriate segmentation parameters within the eCognition software. In this procedure, three parameters (shape, compactness, and scale) are used to guide the segmentation of image-objects. We ran tests for the scale parameter at a rate of 1 to 100 with an interval of 25, and for compactness and shape a rate of 0.1 to 0.9 with intervals of 0.2. We set shape to 0.1, compactness to 0.5, and scale to 50. The multidate segmentation output was evaluated based on visual assessment of segmentation suitability, where the scale parameter influenced directly the size of the objects by generating small and spurious information with low parameter values or merged multiple changed regions with high parameter values; compactness controlled the boundary of segments by generating smooth boundaries with low parameter values or compact with high values; and shape controlled the influence of spectral information on the formation of segments by creating color-influenced objects based on low parameter values or shape-influenced objects based on high values.”
Line 159: Figures (a) and (b). What bands are represented in these illustrations?
Response: Figure 3 displayed the SWIR 2 channel. We addressed the axis label in revision.
Results
Lines 249-253: Would the authors consider merging this one-sentence paragraph with others or expanding it? Perhaps they could explain why they think SWIR 2 band has the best potential to detect deforestation, yet it is not as good in distinguishing between deforestation and seasonal changes?
Response: Addressed in revision.
Line 251: “Although SWIR 2 has the potential to be used to detect deforestation, the Def and Sch distributions for this band resulted in a Jeffries-Matusita separability of 0.92 (scale of 0 – 2), meaning misclassifications between deforestation and seasonal changes were common. Studies have pointed out the limitation of this measure due to the strong seasonality in NDVI forested and non-forested distributions suggesting a spatial normalization to improve JM separability [46].”
Lines 257-260: It is not clear what is meant by 'channels' being 'promisor'. Please rephrase this sentence and if possible, use another term in place of 'promisor'.
Response: Addressed in revision.
Line 257: “For instance, if a forested area is converted to bare soil or a water body, SWIR 2 using the methodology presented here would be suitable for this type of change detection, but if the same area is quickly converted to a crop field or initial forest regeneration, another spectral channel may be more suitable.”
Line 291-293: Not clear why the tasseled cap coefficients used in study were inappropriate. Please elaborate.
Response:
The tasseled cap transformations (TCT) use coefficients derived from a set of sampled locations to convert the original bands of an image into a new set of bands, with defined interpretations that are useful for many applications. Regarding to Landsat 8, Baig et al. (2014) derived tasseled cap transformations based on the follow study areas:
Path/Row | Location | Rationale for selection |
16/34 | West Virginia – USA | Mountainous forests, water bodies and urban features |
16/35 | North Carolina – USA | Agricultural holdings, forest and water bodies |
31/30 | Northern Nebraska, South Dakota – USA | Variety of crop fields, urban areas, airport and bare soil |
39/31 | Idaho – USA | Large water bodies and crop fields |
121/40 | Jiangxi – China | Poyang lake, urban and vegetation areas |
152/41 | Indus River Basin – Pakistan | Indus River, extensive crop fields |
As none of these locations have similarities with Cerrado ecossystem, the poor performance of TCT in our study may be explainded by a weak coefficient relationship.
Lines 305-306: What variation in number of images used is considered 'small', and what is an optimal number of images to include in a change detection?
Response: Regarding to the optimal number of multi-temporal images able to change detection, it is a gap and must be explored in future applications, since a small number of images may not capture all change-objects, and a large number may create an inadequate multi-temporal segmentation by creating small and spurious change-objects.
Line 318: ' …...which have continuous data observations …..' Please rephrase sentence for clarity.
Response: Addressed in revision.
Line 316: “In particular situations, our method may not detect deforestation when it occurs in the last two images and the seasonal noise is still very strong. In this situation, Cerrado trees are mostly non-photosynthetic and the increase in reflectance caused by deforestation may be not be detected as forest change, because another observation must be included to detect the change. In that case, monitoring studies, which continuously acquire observations and imagery, can avoid these undetected change events.”
Lines 321-322: This is a 'one-sentence' paragraph that should either be combined with the previous paragraph or expanded.
Response: Addressed in revision.
Line 321: “In particular situations, our method may not detect deforestation when it occurs in the last two images and the seasonal noise is still very strong. In this situation, Cerrado trees are mostly non-photosynthetic and the increase in reflectance caused by deforestation may be not be detected as forest change, because another observation must be included to detect the change. In that case, monitoring studies, which continuously acquire observations and imagery, can avoid these undetected change events. Thus, this limitation must be explored in the Cerrado, since the difficulty of change detection in this biome is widely.”
Conclusions
Lines 322-325: But you stated earlier that the poorer ability to resulted in higher misclassifications between the two classes. Why the contradiction here?
Response:
We have stated the challenge of detecting deforestation in highly seasonal areas of the Cerrado, and addressed the research question about the possibility of detect land cover change in areas affected by seasonality with high degrees of accuracy. According to our study, we have demonstrated that SWIR 2 band presented high change detection accuracy, differentiating deforestation from seasonal changes, resulting in fewer misclassifications. Thus, we did not contradict ourselves but answered the stated research question.
Line 326: ' …. satisfactory change detection accuracies ….' Addressed in revision.
Author Response File: Author Response.docx
Reviewer 3 Report
The work results very interesting and original. Nevertheless some minor revisions need to be added to reinforce the study before its publication.
First of all, since you considered a very short change period as is from 2013 to 2014 I'm not sure so many changes could happened, so please add more details about this choice. The use of a large time series is declared as one of the limit of the method: what could be the maximum number of multitemporal images able to change detection?
What are the thresholds used in Figure 3 to discriminate between deforestation or seasonal change? and what does a minimum peak could be explained?
The error estimation needs to be added in Table 2 for Overall Accuracy, User's and Producer's according to Olofsson et al. (2014) Remote Sensing of Environment, 148, pp.42-57.
It will be very interesting the testing of the proposed methodology (in particular the SWIR response) for the changes of other different target class (as changes from grassland to cultivated land) at least among future applications as the comparison between the performance of random forest classifier with other ones as for example support vector machine.
Author Response
Response to Reviewer 3 Comments
The work results very interesting and original. Nevertheless some minor revisions need to be added to reinforce the study before its publication.
First of all, since you considered a very short change period as is from 2013 to 2014 I'm not sure so many changes could happened, so please add more details about this choice. The use of a large time series is declared as one of the limit of the method: what could be the maximum number of multitemporal images able to change detection?
Response: Addressed in revision.
Line 100: “We acquired cloud-free Landsat OLI images (path 219, rows 70 and 71) between April 2013 and December 2014 from the United States Geological Survey for Earth Observation and Science (USGS/EROS). The Landsat products were processed at Level-2, supporting time series analyses and data stacking with high precision [14] and bottom of atmosphere reflectance calculated [40]. We set a short two-year period because: (i) we can acquire and process images with lower time gaps and lower computational effort compared to long time series; (ii) the method can detect change events that may not be captured by a biannual method due to a rapid land conversion, for instance; (iii) a short period can generate annual deforestation reports, providing support for continuous monitoring and conservation applications.”
Regarding to the maximum number of multi-temporal images able to change detection, it is a gap and must be explored in future applications, since we have to find an optimal number of images which can capture all changes events, and avoid an inadequate multi-temporal segmentation by creating small and spurious objects.
What are the thresholds used in Figure 3 to discriminate between deforestation or seasonal change? and what does a minimum peak could be explained?
Response: There is no threshold to discriminate Def from Sch. The multi temporal objects were classified based on a set of variables, which are the maximum difference spectral of several spectral indices.
A minimum peak in Gmax is a sudden loss of reflectance values in the original spectral channel signature. This is correlated to the high seasonal variation in the Cerrado vegetation, and could be explained by a quick leaf gain, disturbed by a strong rainfall at some point.
The error estimation needs to be added in Table 2 for Overall Accuracy, User's and Producer's according to Olofsson et al. (2014) Remote Sensing of Environment, 148, pp.42-57.
Response: Addressed in revision.
Line 223: “Table 2. Accuracy analysis results from change detection, represented by overall accuracy and its complementary measure, overall error; producer’s accuracy and its complementary measure omission error (O.E.); and user’s accuracy and its complementary measure commission error (C.E.).”
Site | Overall Accuracy (%) | Overall Error (%) | Producer’s Accuracy (%) | User’s Accuracy (%) | ||||||
Def | O.E. | Sch | O.E. | Def | C.E. | Sch | C.E. | |||
All | 88.3 | 11.7 | 88.0 | 12.0 | 88.6 | 11.4 | 84.6 | 15.4 | 91.2 | 8.8 |
1 | 83.3 | 11.7 | 75.0 | 25.0 | 100.0 | 0.0 | 100.0 | 0.0 | 67.7 | 32.3 |
2 | 87,5 | 12.5 | 100.0 | 0.0 | 75.0 | 25.0 | 57.1 | 42.9 | 100.0 | 0.0 |
3 | 93.3 | 6.7 | 100.0 | 0.0 | 87.5 | 12.5 | 87.6 | 12.4 | 100.0 | 0.0 |
4 | 86.7 | 13.3 | 80.0 | 20.0 | 90.0 | 10.0 | 80.0 | 20.0 | 90.0 | 10.0 |
It will be very interesting the testing of the proposed methodology (in particular the SWIR response) for the changes of other different target class (as changes from grassland to cultivated land) at least among future applications as the comparison between the performance of random forest classifier with other ones as for example support vector machine.
Response: Agreed. We reinforced these ideas by addressing in revision.
Line 330: “Future research should follow-up this approach by exploring different data sources such as Sentinel imagery, along with the use of different classifiers. Further studies should also improve upon the current method, address the issues discussed above, such as the sampling factor, timing of changes, particular change events, and the processing of large spatial (extensive areas) and/or temporal (long time series) datasets. In addition, the response of SWIR 2 to different target classes, such as non-native vegetation, should be investigated further.”
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
The revised manuscript is greatly improved and the authors have done a great job clarifying sections of manuscripts and addressing my concerns. I look forward to seeing the published paper!