CroplandCDNet: Cropland Change Detection Network for Multitemporal Remote Sensing Images Based on Multilayer Feature Transmission Fusion of an Adaptive Receptive Field

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study was aimed at solving poor detection results when employed a change detection method for the dynamic monitoring of cropland. For this purpose, authors proposed a novel cropland change detection network that combines an adaptive receptive field and multiscale feature transmission fusion, and compared with six advanced change detection methods on the CLCD dataset. Although the work claimed that the CroplandCDNet is a novel and effective method for cropland change detection, but the manuscript still needs some revisions before it can be accepted.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This paper reported a change detection network based on Multilayer Feature Transmission Fusion in an adaptive receptive field. The experimental results show improvements with some baselines. Some issues should be carefully addressed before any possible publication.

1. The literature review must be enhanced. A large number of deep learning-based change detection methods are absent, such as doi: 10.1109/TGRS.2022.3157721.

2. Please explain the advantages of the proposed network compared with other Transformer-based or CNN-Transformer-based methods.

3. Please specify the terms of feature transfer/transmit/ transmission layer in the manuscript.

 

4. My biggest concern is about the motivation of the transmission layer. From the eye of the reviewer, the features at multiple levels are fed into the SKA module and transferred to the decoder through skip connections. So the “feature transfer layer” consists of multiple SKA modules and how the transmission through t1_1, t1_2, …, t1_5, which are very confusing for readers. Moreover, Figs 1 and 4 are not informative.

Comments on the Quality of English Language

N/A

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

The authors present a new deep learning approach to detect cropland change in bi-temporal high (circa 1m) resolution satellite imagery. Discussion of multitemporal (not only bitemporal). The model is called CroplandCDNet and it combines an adaptive receptive field and multiscale feature transmission fusion. The model outperforms other state-of-the-art models on a benchmark dataset from China. The work is novel and will be interesting to the audience of Remote Sensing. I can recommend publication of the paper only after the authors address my comments below. I would like to see the revision before it is accepted.

My biggest concern is that the authors have framed the paper too narrowly and there is virtually no discussion section. I would like to see more discussion of how transferable these methods are to medium resolution satellite imagery like Sentinel-2 which are very often used in operational cropland monitoring. For instance, how would the spatial convolutions be adapted given the coarser pixel resolution? How could your model benefit from the multispectral bands in Sentinel-2? How could you leverage the rich temporal dimension of multi-temporal time series as opposed to bi-temporal time series? Addressing such questions would greatly increase the relevance and impact of your work for the broader remote sensing community.

Some smaller comments below:

L35-48: This paragraph has no references. These statements need to be supported with reference to the literature. The sentences are also very long and complicated. I would suggest simplifying. For example “comparison methods based on machine learning and deep learning methods from new machine learning paradigms.” Machine learning is repeated twice in a non-logical way.

L55: Please define “fine cropland change”

L73: “however, post-classification….” Please provide a reference for this? Are there studies that compare post-classification methods with methods based on deep learning? This would be interesting to mention and describe the results.

L75: What do you mean by “construct features manually”?

L87: What are “modules”? Please define.

L106: You cover the application to bitemporal images well. But could you say something about multitemporal (ie. Many images in a time series) deep learning approaches?

L164: Please define what you mean by “high resolution”. I know you mention the spatial resolution further down, but good to state it here as well.

L180: Pleas provide a reference for this sentence.

L305: Is this 0.5 m to 2 m? Or did you mean to write 0.5 meters squared? If the former, please add a space before and after the dash.

 

Comments on the Quality of English Language

The English language is quite good. But many sentences are overly-complex and I would recommend going through and splitting up long sentences or rephrasing.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

Comments and Suggestions for Authors

Comments for author File: Comments.pdf

Comments on the Quality of English Language

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript has been revised and one issue should be further addressed.

As shown in Figure 1, two SKL modules are adopted and each takes multiple features as input and output. However as shown in Figure 4, the network contains multiple SKL modules at different scales. Please revise Figure 1 to clarify the structure of the network.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx