Large-Scale Oil Palm Tree Detection from High-Resolution Satellite Images Using Two-Stage Convolutional Neural Networks
Round 1
Reviewer 1 Report
The paper proposed a method of ML-CNN to detect oil palm trees in high-resolution satellite images with a large-scale. Although the experiment would be practical, I would like to recommend "Reject" because of lacking details of problem definition or research question. Applications of deep learning to remote sensing have appeared quite a lot so far and such "lab-experiments" achieved beyond 90% although those are not always in practice. Authors on this sort of papers should carefully state and define problems to be addressed by proposed methodologies. I would give detail comments in the following.
L100-111 - Please indicate major problems, which are addressed in this paper, for large-scale processing using deep learning. "There is no deep learning based large-scale tree crown detection methods in existing studies." would not enough rationale to accept for publication. L153-161 is likely defining the problem. However, L156-161 is not owing to the large-scale data, just with size of sample images "due to the special pattern or texture of the oil palm plantation areas, it will be much easier to identify the oil palms from other type of vegetation if we utilize the input image that is larger than the size of a single oil palm tree.".
L125-132 - Please indicate observation date and location (e.g., extent or center by latitude and longitude), and DigitalGlobe's image ID if available.
Figure 1. - Please add what green/red/yellow/blue dots represents in the caption, not only in the body text.
Figure 2. - It is not clear how the results in the first, second, and third rows were used together. Because the three are not connected at the end, the results look independent to each other. Although I have reviewed in detail to understand how the land cover classification result was used in the palm tree detection, it is not yet clear for me.
L166 "in three classes" - what are the classes?
L169 "in four classes" - what are the classes?
Figure 5 - Please add labels to the images in the upper column. I suppose the processing of 3.3.3 is between the third and fourth image. If so, please cite Figure 5 in 3.3.3. Labeling the images (a), (b), (c), and (d) would be useful references for the citation in the body text.
Large-scale data processing often concern computing resources and time. Please indicate hardware for the experiments and computing time.
Author Response
Please check the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The authors present a methods for detecting palm trees in high resolution images.
The manuscript is well written and easy to follow. There are minor typos that should be fixed.
Experimental validation is well done. The authors discuss in details the effect of the parameters and the type of the architecture used, and they give sufficient results regarding the performance of their methodology.
I would suggest the authors to clarify the step regarding the fine-tuning of the hyper parameters. Specifically, they must clarify that the test set used for fine-tuning the hyper parameters is not the same with the set used for reporting the performance of their method. Typically a validation set should be used for tuning the hyperparameters. By tuning the hyperparameters on test set, actually the model “overfits” the test set. In real world applications the actual test set is not available, and thus these hyperparameters may not be the best possible for the actual test set. However, since the actual test set is unknown you are not able to report any results. For this reason a validation set (different than the test set) is used for fine-tuning the hyperparameters and then the accuracy is reported using a separate test set.
References are adequate, however, I think that the authors should add these two references
Yushi Chen, Zhouhan Lin, Xing Zhao, Gang Wang, and Yanfeng Gu, “Deep learning-based classification of hy- perspectral data,” Selected Topics in Applied Earth Ob- servations and Remote Sensing, IEEE Journal of, vol. 7, no. 6, pp. 2094–2107, June 2014
And a more recent work
Makantasis, Konstantinos, et al. "Tensor-based classification models for hyperspectral data analysis." IEEE Transactions on Geoscience and Remote Sensing 99 (2018): 1-15.
Author Response
Please check the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
The authors present a multi-level convolutional neural network for oil palm tree detection. The paper is very interesting for practical proposed, but the methodological contribution is very low.
Majors:
1.- I dont understand the title of paper in the sense that CNN models extract multi-level features, from low-level representations at the first layers to more complex and abstract representation at the end of the network.
2.- Figure 5 is very confuse. Is the final predicted label the one obtained by the second model? If yes, then the method proposes two-CNN models: the first one to perform the land cover classification of input patches, and the second one to detect the palms of those patches previously classified as palm plantations. What are the benefits of this strategy in comparison with other object detection methods?
3.- Similar strategy for first CNN is like landcover classification for remote sensing and the second CNN is similar the model publicated by the authors. In fact, the second CNN could be replaced by SVM, MLP, RF… For these reason, in my opinion, the methodological contribution is very low.
4.- In order to make the experimentation to be replicable to the research community, is the dataset avaiable? Please provide the link in the paper. If the dataset is private, they need test the method with a public dataset.
5.- The metrics used on paper: authors need add F1-Score. Moreover, I dont understand the equation of OA proposed by the authors. In my best knowledge, I understand the definition of Overall Accuracy as (Corrected Predicted Class / Total testing class).
6.- I suggest to redo tables 1,3,4,5 and 6. One table by region with all methods, using bold for emphasizing better result of the row. I think this will be better to underestand the performance. Finally, remove table 6 if the authors consider very redundant (I think so, but it is ok also). Also, add ANN.
7.- The performance of the proposed “Deep Learning Based Oil Palm Tree Detection and Counting for High-Resolution Remote Sensing Images” is very bad now. I think that the authors has a lot of overfitting, and they need insert some mechanism to remove false positives (First CNN of this paper).
Minor:
1.- Figure 4: Improve the visualization of the images.
p { margin-bottom: 0.25cm; line-height: 115%; }
Author Response
Please check the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
I confirmed my comments are addressed in the revision.
