NRN-RSSEG: A Deep Neural Network Model for Combating Label Noise in Semantic Segmentation of Remote Sensing Images
Round 1
Reviewer 1 Report
Please see it in the attachment.
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The paper proposes a loss function using reverse cross-entropy on remote sensing image segmentation. This is a well-known trick in machine learning literature.
The term NRN noise robust network does not reflect your work. Your proposal is not a general proposal for noise. The proposal is an application of reverse cross entropy designed for image segmentation. Change the proposal to a name suitable for what you are doing.
Lines 255 "These were used to simulate the symmetric and asymmetric label noise introduced in the annotation process, and also tested on a clean sample set". Please add more info about the proportion of symmetric and asymmetric in your result. Also, it would be nice to have an evaluation to evaluate the sensitivity of the proposal with different ratios of symmetric and asymmetric noise.
My suggestion to the authors is to perform an ablation study showing a stacked result of the RCE, RCE + CBAM on the test set.
Also, it is crucial to run a significance test comparing your proposal with the baseline.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
I have found this manuscript clear and well written. Its content provides relevant ideas to contribute in an interesting way to the reduction of the impact of mislabeling on remote sensing image segmentation.
The Authors have provided a well-structured exposition of their material with a gradual description of the underlying ideas fully appreciable. The content is self-explanatory and described with a sufficient level of detail to understand the topic, techniques and results.
1) The experimental part is however limited and offers only a first illustration of the efficiency of the proposed method on a single HSI data set. In other words, the content is currently limited to a proof of concept.
Yet, the analysis provided is well conducted and corresponding results are fully appropriate to the text and its content. The list of references to the literature related to the field is also quite appropriate.
Here are some extra observations to reinforce the weak points and in particular to extend the discussion preceding the conclusion:
2) The Authors should avoid giving too much numerical detail in the Abstract and instead report the essential trends in a more synthetic way.
3) One question concerns the setting of the two hyperparameters: alpha and beta. How do you proceed in a real-world situation to find an appropriate value for these two parameters, just from the knowledge of the available training data?
3) I suggest adding practical recommendations if possible, especially since the best experimental values obtained here on the Vaihingen dataset seem to depend on the noise rate (unknown a priori)?
4) Can the settings of these hyperparameter values made for this data set be generalized to other data sets?
5) Another avenue of discussion concerns the generalizability of what is proposed here to other semantic segmentation models. Please discuss further
Typo :
the modified loss function is presented in.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Dear Authors
I have now completed the review of the manuscript titled “NRN: A Deep Neural Network Model for Combating Label Noise in Semantic Segmentation of Remote Sensing Images”. This study examines the semantic segmentation of remote-sensing images that include label noise. It proposes an anti-label-noise network framework, termed noise robust network (NRN) to combat label noise. The algorithm combines three main components: network, attention mechanism, and a noise-robust loss function. The performance of the network was evaluated by comparing the NRN with the original U-Net model.
The topic is quite interesting and relevant. I have few comments to improve the quality and clarity of the manuscript.
- The results discussed in the abstract should be reduced in size for better understanding.
- In Figure 1., the clutter class is not shown in the labelled image.
- Line 39-41: With the rapid development…images, I suggest adding [1].
- Figure 4. B, please add the label for the layers after sharing MLP.
- Figure 4. B, please add why the authors choose maxpool and average pool, why not both max or average pool?
- Line 248, please add the reference for FWIoU.
- Authors should add the computational complexity of the RF model, see and add CDLSTM, SMOTEDNN, etc.
- Results and discussion are represented in a detailed manner.
[1] CNN Based Automated Weed Detection System Using UAV Imagery
[2] SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 5 Report
The authors examined the semantic segmentation of remote sensing images that include label noise and proposes an anti-label-noise network framework, termed noise robust network (NRN) to combat label noise. The paper is well-structured, but the language must be refined. The introduction is well-written, but the methods must include the assessment of statistical methods. The results must be detailed. Also, the results are not discussed with the literature.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors fullfiled most of my concerns.
Reviewer 3 Report
The Authors have generally answered correctly the questions I had previously raised.
The quality of the content of the manuscript has been improved.
The current content allows to get an objective idea of the strengths, weaknesses and limitations of the proposed method.
The revised content appears to me to be acceptable for publication.
Reviewer 5 Report
The manuscript is improved, and it can be accepted
