Exploring Transfer Learning for Anthropogenic Geomorphic Feature Extraction from Land Surface Parameters Using UNet

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study presents a valuable exploration of transfer learning for anthropogenic landform feature extraction, specifically focusing on terrace edge detection and valley fill slopes. The key findings are interesting and valuable. However, there are areas where further clarification and refinement are needed.

 

1.     The study focuses on anthropogenic landforms, a unique type of geomorphic feature on the Earth’s surface. However, the introduction lacks sufficient detail on anthropogenic landforms and the rationale for selecting them as the study’s focus. I suggest adding a section that explains the reasons for choosing anthropogenic landforms and references relevant studies. Additionally, there is some redundancy in Section 2, especially regarding the descriptions of CNN and UNet, which could be streamlined.

2.     The terrace labeling and task in this study focus on terrace edge detection, which differs from traditional terrace mapping that usually includes both terrace surfaces and edges. I recommend clarifying the distinction between terrace edge detection and traditional terrace mapping, along with an explanation of this points.

3.     The use of geomorphons as labels is a good choice, yet different window sizes in various areas may affect the results. This point worth consideration, as geomorphons are a primary data source for the transfer task. Discussing the potential effects of varying window sizes on the results would strengthen the study’s analysis.

4.     I noticed that the loss and F1-score curves for the frozen and unfrozen models in the terrace task are similar, leading to less distinguishable figures at the end of the epochs. Consider enhancing the figures to more clearly display the differences between the models over the training epochs.

5.     This study demonstrates the effect of sample size in training models, which is valuable; however, Sections 4.3 and 4.4 also consider this impact into result. Therefore, Section 4.2 may not be essential or could be integrated into the Discussion section.

6.     Although the study focuses on transfer learning performance in geomorphic mapping, the UNet architecture is somewhat outdated. It may be beneficial to consider and compare newer architectures, such as GANs or Transformers, to enhance the study’s relevance and potential impact.

7.     The finding that ImageNet transfer learning performs better than other geomorphic models is interesting and unexpected. I recommend including this important insight in the abstract, as it could inform the design of foundational models for deep learning-based geomorphometry. Additionally, it would be beneficial to analyze if this result is influenced by specific terrain features. If feasible, consider examining additional features (e.g., slope, aspect, and relief).

Author Response

See attached.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

I recommend its acceptance for publication in Remote Sensing as it presents valuable advancements for the field of geomorphic feature extraction, addressing a challenge of labeled training data. The authors examine how transfer learning from large, pre-existing datasets like ImageNet may reduce the reliance on extensive, domain-specific labeled data. The study’s methodology is well-designed and rigorous, exploring two significant geomorphic extraction tasks: agricultural terraces and valley fill faces, using LSPs derived from LiDAR-based DTMs. By comparing the effectiveness of different transfer learning approaches, including ImageNet and geomorphon-based parameters, the authors provide a comprehensive view of transfer learning usage and performance in the model. The findings that ImageNet-based parameters enhance extraction accuracy are particularly noteworthy, as they suggest a broader application of generalized image datasets in remote sensing. While the results indicate limitations in transfer learning when working with smaller sample sizes, this underscores the importance of ongoing research into unsupervised and semi-supervised learning methods. The authors’ discussion of these challenges is transparent and points toward meaningful directions for future studies, making this work a relevant and timely contribution to the field. I believe this manuscript provides practical recommendations for remote sensing practitioners.

Author Response

See attached.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript is about Exploring Transfer Learning for Anthropogenic Geomorphic 2 Feature Extraction from Land Surface Parameters using UNet.

The manuscript and its structure have been reviewed and meet the journal's standards.

The manuscript is well structured, and there is a use of citations.

Author Response

See attached.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

Dear Authors,

Congratulations on the work and thank you for the opportunity to review your manuscript. The article is well-written, organized, and clearly presents its objectives, methodology, results, and conclusions. It demonstrates meaningful incremental contributions to the ongoing research in mapping Anthropocene geomorphology. The discussion is thorough, and the integration of findings within the context of existing literature is notable.

I also appreciate the balanced discussion of the study's limitations and your suggestions for future research. 

Overall, this is a solid contribution, for which I have only minor comments:

Fig 1 (a): If you have training, validation and test data in (b), shouldn't you have 3DEP tiles in Iowa?

lines 162-164: remove parenthesis 

line 321: valley fill faces is in (c), not (b)

 

Author Response

See attached.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript is well organized and clearly outlines the goals, methods, and results. The findings are interesting and valuable for DL-based geomorphometry. I suggest to accept in this version.