On-Board Volcanic Eruption Detection through CNNs and Satellite Multispectral Imagery

Round 1

Reviewer 1 Report

1. Line 16. Must add more RS applications, e.g.,

[1] Li, J.; Qu, C.; Shao, J., Ship detection in SAR images based on an improved faster R-CNN. In SAR in Big Data Era: Models, Methods and Applications (BIGSARDATA), 2017; pp 1-6.

[2] Zhang, T.; et al., HOG-ShipCLSNet: A Novel Deep Learning Network With HOG Feature Fusion for SAR Ship Classification. IEEE Trans. Geosci. Remote. Sens. 2021, early access, 1-22.

[3] Bentes, C.; Frost, A.; Velotto, D.; Tings, B. In Ship-Iceberg Discrimination with Convolutional Neural Networks in High Resolution SAR Images, Proceedings of EUSAR 2016: 11th European Conference on Synthetic Aperture Radar, 6-9 June 2016; 2016; pp 1-4.

[4] Mao, Y.; Yang, Y.; Ma, Z.; Li, M.; Su, H.; Zhang, J., Efficient Low-Cost Ship Detection for SAR Imagery Based on Simplified U-Net. IEEE Access 2020, 8, 69742-69753.

2. Line 19. Consider the following to confirm your insights, e.g.,

[1] LeCun, Y.; Bengio, Y.; Hinton, G., Deep learning. Nature 2015, 521 (7553), 436-444.

[2] Wu, X.; Sahoo, D.; Hoi, S. C. H., Recent advances in deep learning for object detection. Neurocomputing 2020, 396 (2020), 39-64.

[3] Mao, Y.; Li, X.; Su, H.; Zhou, Y.; Li, J. In Ship Detection for SAR Imagery Based on Deep Learning: A Benchmark, 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), 11-13 Dec. 2020; 2020; pp 1934-1940.

3. Line 22. Have the authors tried the on board training and test? What is the training and test time? How about the model size? It your model lightweight? Must consider shipdenet-20: an only 20 convolution layers and <1-mb lightweight sar ship detector in your discussion.
4. Line 29. Which type of satellite? SAR or optical? Similar must-be consideration from Sentinel-1 like LS-SSDD-v1.0 dedicated to small ship detection?
5. Line 38. “from classical to AI techniques as for example in [10]”. The [10] has similar work as yours. What is the difference between yours?
6. Line 4. “a first prototype for an Artificial Intelligence (AI) model to be”. First one?
7. Line 44. What is your on-board system? Jextson? See:

[1] Xu, P.; Li, Q.; Zhang, B.; Wu, F.; Zhao, K.; Du, X.; Yang, C.; Zhong, R., On-Board Real-Time Ship Detection in HISEA-1 SAR Images Based on CFAR and Lightweight Deep Learning. Remote Sens. 2021, 13 (10), 1995.

8. Line 56. Can you make this dataset open to everyone? Like LS-SSDD-v1.0?
9. Line 138. Please consider high-speed ship detection in sar images based on a grid convolutional neural network and depthwise separable convolution neural network for high-speed sar ship detection in classification task [19–21].
10. Figure 4. The output should be two-dimension, rather than the only one. You can refer to hog-shipclsnet.
11. Give a reference model from the CV community about Figure 4.
12. Line 154. Give more details about the balance operation of your dataset. Rotation operations in balance scene learning mechanism for offshore and inshore ship detection in sar images?
13. Figure 5. What is the accuracy? The classification accuracy of training set and validation set? I have not seen their definition.
14. Line 181. Just remove more layers? Have you considered depthwise separable convolution in depthwise separable convolution neural network for high-speed sar ship detection?
15. Line 200. This section is rather interesting. Please expand it and add more details.
16. Table 8. I think that the speed is still slow on the strawberry system when compared with hyperli-net: a hyper-light deep learning network for high-accurate and high-speed ship detection from synthetic aperture radar imagery. You can add some texts, e.g., CPU, GPU, MCU, ARM?
17. This work is lack of some method comparison. The authors can consider add more backbone networks to confirm the real-time advantage of your model, e.g., ResNet, DenseNet, and so on. Tens of must-be competitive comparisons in quad-fpn: a novel quad feature pyramid network for sar ship detection. Please consider.

Carefully solve the above problem for possible accept.

Author Response

Please see the attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors present an interesting case study on the application of AI to detection of volcano eruptions in multispectral images through on-board processing in spaceborne platforms. The manuscript starts from an introduction to the problen, then describes how an extensive dataset including several cases of eruption was built by manually labeling a set of multispectral satellite images. Then the manuscript describes the model and the training strategy, to then discuss how it was simplified to fit on a simpler hardware simulating typical on-board processing capabilities. Performances are then assessed for both the original and the simplified model, and results discussed.

The authors appear to have done a great deal of work in preparing and curating the data used in the experiments, as well as in the practical realization of the simulation experiments. Practical results seem convincing, and I can recommend publication of the manuscript after some minor issues have been resolved:
- in section 3.1, the authors mention they used a library from the Phi-Lab to tackle the problem of the severely imbalanced training set. However, the explanation of how the problem is solved is unclear. Given the importance of the problem, I would provide the reader some more information on the subject;
- in section 5.1, the authors write of "random prediction", probably meaning inconclusive classification results. The term "random" does not seem appropriate to me in this case. I suggest to use a more descriptive one.
- in section 5.1 and 5.2, the set of results is rather small, and performances are not summarized into a single figure (like e.g. accuracy). The authors should at least explain why they consider the small test set representative and possibly compute classification accuracy figures.

Author Response

Please see the attached file.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

All concerns have been solved.