An Encoder–Decoder with a Residual Network for Fusing Hyperspectral and Panchromatic Remote Sensing Images
Round 1
Reviewer 1 Report
The paper develops a data fusion technique for constructing high-spatial-resolution hyperspectral images from (low-resolution) hyperspectral images and (high-resolution) pan-chromatic images. The proposed approach is based upon deep learning models. The paper also compares the proposed method with other techniques.As written, the paper suffers from several shortcomings.
1. Fig. 3, 4 and 5 are unclear. What does an "RGB 3D Cube of Hyperspectral image" mean? This description is very confusing. 2. How metrics (results section) are computed exactly. Specifically, how are these values computed for pixels for which no ground truth (in high-resolution) is available? It is worthwhile to expand this section and describe exactly how these results are computed---are these pixel based, layer based, or defined over the entire datacube. 3. It is also not clear how the final datacube is constructed. Does the method use mosaic techniques (per layer) and then stack layers to construct the high-resolution hyperspectral datacube. 4. Details on model training needs to be expanded. As it stands, it is not clear at all how the model was trained. Does it treat each layer differently? How does one decide upon a batch size? What training platforms were used? What sort of computational resources are needed to train such models? 5. The paper should include some discussion of the upsampling step. Upsampling step can be very costly for hyperspectral images with a large number of bands? 6. On a more philosophical note, while it is great to treat upsampling in isolation. Oftentimes upsampling is performed to carry out a subsequent task, such as classification or segmentation. It will be useful to include some discussion about the how upsampling might affect a downstream task. It is possible that there sia sweet spot for upsampling and if the image is upsampled beyond this point, the accuracy/performance of the downstream task begins to diminish. 7. The write-up can also be improved. The paper includes some very long sentences and some very long paragraphs, which adversally affects the overall readibility of this paper. The paper also makes extensive use of passive voice. It may improve readibility to write short sentences in active voice. 8. Along the same lines, I am not sure if the paper needs to discuss deep learning background, which is widely available at other places. The paper can make better use of this space by including more details about the 1) implementation, 2) results, and 3) limits.
In order to ensure the reproducibility of this method, I will strongly encourage the authors to release the benchmarks and source code for this work.
On a positive note, however, the work carried out in this work is timely and highly relevant or the readers of this journal.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 2 Report
Dear authors,
I found another publication https://arxiv.org/pdf/2107.02630.pdf
Hyperspectral Pansharpening Based on Improved Deep Image Prior and Residual Reconstruction
I didn't go in too much detail, but it seems the authors conducted a similar work. Could you please explain how your work differs from theirs? Or what additional contributions you provide in yours?
Regarding your work, the datasets should be further described. What are you considering as ground truth? How did you split your data (train, validation and test)? What about the model? What batch size did you use? What image input the network had? Did you use a pre-trained model? Loss function? Optimizer? I would also add more qualitative images, zoom in to check small details.
What about methods HPF and SFIM? It seems they are not properly introduced. Figure 8, in the caption, your method should use letter h) instead of g)
In order to consider your work for publication, a clear contribution w.r.t. the work, and other similar ones described in there, I shared should be included.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
Dear authors,
I see you have definitely improved your paper. Despite the contributions you claim in your work wrt the work I mentioned in my previous review, I do think a comparison between both methods should be conducted. Thankfully, the authors from the paper provide code and links to the datasets they used to validate their method:
https://github.com/wgcban/DIP-HyperKite
You could simply run your method on these scenarios they share, and add these results to your work. This way, we will be able to compare both methods in the same table.
If you include these results I would recommend your work for acceptance.
Kind regards,
Author Response
Please see the attachment.
Author Response File: Author Response.pdf