Polarimetric Synthetic Aperture Radar Image Classification Based on Double-Channel Convolution Network and Edge-Preserving Markov Random Field
, Mengmeng Nie 1,2, Shanshan Ji 1,2, Cheng Shi 1,2, Hongying Liu 3 and Haiyan Jin 1,2,*Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsIn this paper, the authors apply a two-level feature-fusion model based on double-channel convolution network and edge-preserving Markov random field, and constructs a PolSAR image classifier to demonstrate the effectiveness of the proposed model. Generally, this paper is well organized, but it seems a combination of existing methods like DCCNN and MRF. The novelty of this manuscript should be further highlighted.
Detailed comments are:
1) The image quality of Figure 2 requires enhancement.
2) In Section 3.1 3), the authors have mentioned that the proposed DCCNN fusion network can suppress the useless feature. Please explain how it works because the authors simply engage the known PolSAR data decompositions without comparing ones informativeness and correlatability. An eclectic set of unreasonable identification features does not guarantee good classification quality.
3) In Section 3 Algorithm 1, the authors generate the class probability and estimated class label map by the DCCNN model. The probability obtained by the Softmax layer is biased, which could potentially interfere with the MRF reasoning process. Therefore, probability calibration is necessary.
4) In Section 4.3, why is the overall accuracy of CV-CNN in this paper 2% higher than the original model used in Reference 44? Furthermore, in this paper, CV-CNN demonstrates 100% accuracy for buildings with complex scattering mechanisms, whereas in Reference 44, it is only 83.2%. I suggest that the authors clarify the reasons for these disparities.
5) What is the noise level of the dataset? Any despeckling method applied before training? Please explain it more.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper “Polarimetric SAR Image Classification Based on Double-Channel Convolution Network and Edge-preserving Markov Random Field” details an algorithm that uses a double-channel CNN network combined with an edge-preserving MRF model to improve PolSAR image classification. The authors introduce a new method called DCCNN-MRF, which integrates the edge-preserving MRF with the DCCNN model. This combination aims to reduce speckle noises and refine the image edges. The method was tested on four distinct sites with data from RADARSAT-2 and AIRSAR. These sites, which include rivers, buildings, and crop areas, provide a comprehensive testing ground. For comparison, the authors used several methods: Super-RF, DBDA, S3ANet, CV-CNN, DCCNN, and DCCNN-MRF, with DCCNN-MRF showing superior accuracy. The authors also examined the effect of patch size on accuracy, analyzed the training sampling ratio, and assessed the running time. In conclusion, DCCNN-MRF required only 10% of the training samples, a rate slightly high for deep learning methods.
For the authors: Section 4.1 lacks coherence in its figures and is challenging to follow. Please revise the sentence on pages 409-411. There are no discussions/comparisons with other studies that try to improve PolSAR image classification.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsIn general, the technical structure of the paper is well done. The idea although, it is a technical procedure is very well documented and can be considered novel.
The paper is well supported by mathematics.
However,
1. The authors should read the manuscript carefully and rewrite the abbreviations after the first appearance of their meaning. It is better in the abstract to give the full name.
2. Some recent references are needed (state of the art):
a. A New Architecture of a Complex-Valued Convolutional Neural Network for PolSAR Image Classification
b. Simulated Annealing for Land Cover Classification in PolSAR Images
c. Fully Polarimetric Land Cover Classification Based on Hidden Markov Models Trained with Multiple Observations
d. Fully Polarimetric Land Cover Classification Based on Markov Chains
3. The authors should state the novelty of their work more clearly and what are its advantages and disadvantages compared to relative works.
4. The authors should give a block diagram with the steps of their algorithm before the description of the algorithm.
5. Finally, the authors should give in the conclusion extended the highlights for future research and give more explanation for the results of their work.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
