Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors present a new deep learning pipeline that uses attention-based deep learning to perform state-of-the-art semantic classification of fluvial landscapes with Sentinel 2 imagery accessed via GEE. They train, validate and test the network on a multi-seasonal and multi-annual dataset drawn from a study site that covers 89% of the Earth’s surface. The authors have done much work, but the manuscript is very poor now, it is more like a report.

 

(1) The authors should rewrite the manuscript base on the Journal requirement, now the format is coarse.

(2) Where are new things in this research? Method? OR just efficiency?

(3) The validation is not enough! The study area is almost the whole earth.

(4) The details of the models applied should be given.

(5) There are many mistakes in this research. Line 162; Figure 4,5,6, the legend is lost. Line 163-173; Table 1 is very poor.

Author Response

Please see attached response letter to review 1.

 

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

In this paper, the authors introduce a deep learning pipeline designed to perform semantic classification of rivers as a distinct category from lakes using Sentinel-2 data. The manuscript is well-structured; however, several issues warrant attention in this review round.

Lack of Justification for Model Selection:

The introduction presents the choice of the attention-UNet and Segformer Vision Transformer architectures without fully explaining why these models were specifically selected. While they are popular in the field, the text doesn’t clarify what unique benefits these models bring to the study’s context. A stronger introduction should ground the model choice in specific performance or application-based advantages, especially if it’s a standard approach that may lack novelty.

Vague Problem Statement:

Although the introduction highlights that rivers are ecologically crucial and occupy less than 1% of the Earth's surface, it lacks a compelling narrative around why semantic segmentation at this scale and with this approach is significant. For instance, there’s little explanation about existing challenges that previous methods face, such as computational limitations, that this work purportedly addresses. The lack of depth in the problem statement weakens the case for the manuscript’s relevance and innovation.

Standardized Model Approach Without Clear Innovation:

The manuscript's methodology follows a well-established pipeline using a combination of attention-UNet and Segformer Vision Transformer, but it lacks a distinct innovation or novel adaptation to differentiate it from other works. The approach seems to replicate existing models rather than proposing a fresh perspective or novel methodology, which is critical for a publication focused on deep learning advancements.

Unclear Rationale for Fusion Mechanism:

The manuscript mentions a “novel straightforward architecture” for fusing outputs from the models but does not provide enough context or experimentation to explain why this fusion method was chosen. For instance, line 138 refers to an added “extra skip and concatenation connection” for learning fine-grained details, but no comparison is provided to alternative fusion strategies. Without empirical evidence or a theoretical basis, the chosen fusion mechanism appears arbitrary, which diminishes the methodological rigor.

Overuse of Implementation Detail:

Several sections, as previously noted, delve deeply into implementation specifics (e.g., pre-trained TensorFlow B3 model in line 123, downsampling techniques, and augmentation parameters). While some level of technical detail is necessary, the balance here leans too heavily toward explaining basic setup rather than conveying methodological insights.

Insufficient Comparative Analysis:

The results section presents F1 scores and examples from both seen and unseen datasets, yet fails to include any direct comparison to competing models or baselines. For example, although previous works by Carbonneau and Bizzi (2024) and Nyberg et al. (2023) are mentioned, there is no quantitative comparison or discussion to illustrate how this study’s approach improves upon these methods. Such comparisons are crucial to validate claims of enhanced performance.

Overemphasis on Qualitative Results:

The manuscript includes several figures (e.g., Figures 5 and 6) showing sample segmentations but does not provide sufficient statistical analysis or robust quantitative metrics to substantiate its claims. A strong results section should focus on statistically significant improvements, error rates, or confidence intervals rather than relying heavily on visual output alone.

Lack of Novelty in Model and Fusion Technique:

Combining attention-UNet with Segformer Vision Transformer has become a common practice in segmentation tasks, and the manuscript lacks a clear innovative aspect that distinguishes it from other similar studies. The fusion mechanism used is described as novel but is not sufficiently supported or differentiated, making it difficult to perceive this as a significant contribution.

Missed Opportunities for Advancing the Field:

The work could benefit from exploring novel architectures, unique data handling methods, or even innovative applications of the model specifically tailored to fluvial landscapes. As presented, the study does not seem to advance the current state of knowledge meaningfully, nor does it introduce any transformative techniques that would move the field forward.

Comments on the Quality of English Language

A lot of writing issues are detected, indicating that the manuscript lacks editorial precision and clarity, which can significantly impair reader comprehension and engagement. Without a thorough reorganization to remove redundancy, unify pronoun usage, and focus on the essentials of methodology and findings, the manuscript does not meet the standards for readability and professionalism expected in academic publishing. Specifical cases:

Line 45: "I leverage two established architectures, the attention-UNet (Oktay et al., 2018) and the Segformer Vision Transformer (Xie et al., 2021)." Here, the first person singular "I" is used to describe the author's actions.

Line 65: "I start by using the classification outputs from Carbonneau and Bizzi (2022, 2024)." The use of "I" again makes it seem like a single-authored, informal report rather than an academic manuscript.

Line 123: The pronoun usage switches to "we", stating, "We use the TensorFlow (Abadi et al., 2016) implementation of the B3 variant, pre-trained with 1.2 million samples in the ImageNet-1K dataset (Deng et al., 2009)." This inconsistency between "I" and "we" can be jarring to the reader, as it does not establish a unified voice.

Line 70: "Google Earth Engine (GEE) was used via the Python API to download bands 8, 4, and 3 of Sentinel-2 imagery for composited for an entire month." While it’s relevant that GEE was utilized, detailing the specific API (Python) and the process used for downloading can be summarized. This level of detail might interest a developer but detracts from the scientific focus and feels extraneous in the main narrative.

Line 115: The author specifies, "begin with code from Tomar (2024) modified to 1) include a fourth encoder/decoder pair." The mention of this modification would be more concise without the specific reference to code provenance, which doesn’t add value to the methodology.

Line 123: "We use the TensorFlow (Abadi et al., 2016) implementation of the B3 variant, pre-trained with 1.2 million samples in the ImageNet-1K dataset (Deng et al., 2009), available in the Hugging Face Transformers Python package (Hugging Face, 2024)." This sentence contains an overload of implementation details, such as the library and dataset specifics, which could be moved to an appendix or supplementary materials if deemed necessary. Focusing on the function and role of TensorFlow in the research would enhance readability without sacrificing information.

Author Response

Please see attached response letter to review 1

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The author has improve the manuscript totally.

Author Response

We thank the reviewer for this endorsement of our paper.

Reviewer 2 Report

Comments and Suggestions for Authors

I appreciate the author's effort in addressing my concerns in the last revision round. However, the current manuscript is not sufficiently ready for publication due to several deficiencies in literature investigation and contextual framing.

1. Insufficient Literature Review

The author’s claim on pages 2, 3, and 6 of their response—that they "find no citations where this combination of specific architectures (attention UNets and Segformer ViT) was used for land-use classification"—is inaccurate. Several relevant studies utilize similar combinations or architectures for analogous tasks in land-use or environmental classification. These include:

• Zhou, Nan, et al., ViT-UNet: A Vision Transformer Based UNet Model for Coastal Wetland Classification Based on High Spatial Resolution Imagery (IEEE JSTARS, 2024).
• Tong, Qixiang, et al., STIRUnet: SwinTransformer and Inverted Residual Convolution Embedding in UNet for Sea–Land Segmentation (Journal of Environmental Management, 2024).
• Zhao, Xili, et al., A Method for Extracting Lake Water Using ViTenc-UNet: Taking Typical Lakes on the Qinghai-Tibet Plateau as Examples (Remote Sensing, 2023).

It's ok that they missed several pieces of literature in the first submission. However, the insistence on ignorance of such existing literature in this revision round weakens the manuscript’s novelty claim. A more thorough investigation into related works would have helped contextualize the study's contribution.

2. Novelty of Fusion Methodology

The assertion that the proposed dual-input fusion approach is novel also lacks sufficient justification. While the authors suggest that their fusion strategy is unique for combining outputs of two models, similar image-fusion techniques have been explored, albeit in varying contexts. Their request for citations from the reviewer further indicates a limited exploration of the broader literature base.

3. Overlooked Contributions in Related Fields

Despite the claim of insufficiently relevant precedents in land-use classification, methodologies combining Vision Transformers with UNets have been applied effectively in domains with significant overlap, such as wetland classification and sea-land segmentation. These domains share methodological parallels with land-use classification and challenge the claim of distinctiveness.

Author Response

Please see attached letter.

Author Response File: Author Response.pdf