Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. The introduction is somewhat lengthy, with considerable repetition of background information. It is recommended to streamline the content, focusing on the difficulties of decoration waste management and the innovation of this research method, highlighting the unique value of "remote sensing + neural networks" in precise control.

2. Although the remote sensing image recognition process used a combination of manual and machine learning methods to extract decoration waste dumps, quantitative accuracy metrics for the recognition results (such as Kappa coefficient, F1-score, etc.) were not provided. It is recommended to add accuracy validation to enhance the credibility of the remote sensing analysis.

3. The BP neural network model only uses data from the previous four consecutive years as input, without explaining the rationale for selecting this time window. It also does not consider the potential impact of external factors such as policies, economy, and population on the amount of decoration waste generated. It is recommended to supplement the basis for variable selection and discuss the possibility of introducing more dimensional features.

4. Although the model has a high goodness of fit, it lacks sensitivity analysis or confidence interval estimation for the prediction results, which weakens the explanation of result robustness. It is recommended to add a discussion on uncertainty to provide a more comprehensive assessment of prediction risks.

5. Some figures (such as Figure 5 and Figure 10) have deficiencies in legends, color contrast, and label clarity, which affects reader comprehension. It is recommended to redraw these charts to ensure clear legends, distinct color differentiation, and standardized labeling formats.

6. The language expression is occasionally repetitive and lacks fluency. For example, the term "decoration waste" is repeated with high frequency, and some sentence structures are overly long. Language polishing is suggested to improve conciseness and readability.

7. The discussion on the reasons for prediction fluctuations and regional differences could be further deepened. The paper mentions the impact of the epidemic on the 2020 prediction error, but does not systematically analyze the interference of external factors on model performance. The differences in output between cities are only attributed to economy and population, lacking deeper mechanistic exploration. It is recommended to deepen the discussion to enhance the policy guidance and academic depth of the conclusions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Major comments:  

A) Novelty and contribution need to be clarified / The authors should explicitly answer  

1) What is the methodological novelty?

2) Is the novelty mainly the regional case study, the integration framework, the dataset, or the control strategy?

3) How does this study improve on prior work that already used remote sensing or machine learning for construction waste analysis?   Without a clearer statement of contribution, the paper currently reads more like an application study than a strong methodological advancement.  

B) The remote sensing interpretation workflow is not described in enough detail  

1) What exact machine learning method was used in image recognition?

2) What features or bands were used from the GF-2 imagery?

3) How were the six waste categories operationally distinguished?

4) Was there ground-truth labeling or field verification for the identified waste piles?

5) What was the classification accuracy, confusion matrix, precision, recall, or kappa coefficient?   This is a major weakness because the remote sensing interpretation is one of the two central pillars of the paper. If the identification stage is uncertain, then the later conclusions about the spatial distribution and decoration waste share are also uncertain. The paper reports percentages for Pingdingshan, Xuchang, and Shangqiu, but it does not provide a formal accuracy assessment for the classification procedure.  

C) The BP neural network model is under-explained and weakly validated  

The manuscript describes a BP neural network with 4 input nodes, 1 output node, and four hidden layers with 4, 5, 5, and 5 nodes, selected based on “experience” and best fit. The model was trained with 80% of the data and tested on 20%, with a maximum of 1000 epochs and target error rate of 0.00001. The reported overall fit is R = 0.95463 and the prediction error for 2010–2021 is below 15% however:

1) The dataset appears very small for a neural network model. The time series spans 2006–2021, which is limited, and it is unclear how many effective training samples were generated.

2) The authors do not explain why a neural network is preferable to simpler time-series or regression approaches given the small sample size.

3) There is no comparison against benchmark models such as linear regression, ARIMA, random forest, gray model, or LSTM.

4) The manuscript reports mainly R and relative error, but does not provide other standard metrics such as MAE, RMSE, or MAPE in a formal table.

5) It is not clear whether the validation is truly out-of-sample in a time-aware sense, which is especially important for forecasting problems.  

D) The paper would be stronger if the authors discussed:  

1) the relationship between urban development intensity and decoration waste

2) whether policy differences across cities may affect waste generation or disposal

3) whether image-detected dumps represent only unmanaged sites rather than total waste generation

4) How the forecasting results could realistically support planning, monitoring, or intervention.   Overall Suggestions for improvement  

1) Clearly state the novelty and the main scientific contribution of the paper.

2) Expand the remote sensing methodology section with full technical details and classification validation metrics.

3) Provide a more rigorous description and justification of the BP neural network architecture.

4) Compare the BP model with at least one or two baseline forecasting methods.

5) Add uncertainty analysis or sensitivity discussion, especially for the 0.1 t/m² coefficient and future forecasts.

6) Strengthen the discussion by connecting results to urban policy, planning, and sustainability implications.

7) Add a dedicated limitations section.

 

Comments on the Quality of English Language

 

• English language needs significant editing for grammar and clarity.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Manuscript ID:  sustainability-4217327

The manuscript presents a relevant case study on decoration waste management in Henan Province, combining remote sensing interpretation with a BP neural network for prediction. Despite the effort done by the authors some comments should be taken in consideration.

1. The abstract states that high-production areas are distributed in the "southwestern and southeastern" Henan Province, but then lists key cities including Zhengzhou, which is in north-central Henan. This geographic contradiction needs to be resolved.
2. For the keywords, it is recommended to replace "precise control" with a more specific term, such as "waste management" or "generation prediction," to better reflect the study's core contribution.
3. In Section 2.2, the authors mention a combination of manual visual interpretation and machine learning for extracting waste distribution. Please specify which machine learning algorithms were used for the extraction process and provide the classification accuracy metrics.
4. The manuscript uses the "generation per unit area" method (0.1 t/m²) with data from the Henan Provincial Statistical Yearbook to train the BP neural network. However, the remote sensing imagery (GF-2) is used to determine that decoration waste accounts for approximately 10% of construction waste piles. It is unclear how these two independent datasets are integrated. If the BP network's training data is derived from the same statistical yearbook used to validate its predictions, the validation is not independent.
5. The manuscript claims to form a "precise control technology system" and a "foundation for supervision." However, the methodology is a straightforward application of existing remote sensing interpretation and a standard BP neural network. The combination is not inherently novel, and the paper does not demonstrate how this system achieves "precise control" beyond prediction and mapping.
6. Line 30: "from the beginning to the end of the distribution of multi-points to show the characteristics" is a fragmented and unclear sentence.
7. Line 186: "people's daily life is closely related to the arbitrary pile not only affects the city appearance" is a grammatically incorrect and confusing sentence.
8. The manuscript states the network uses "four hidden layers, with nodes of 4, 5, 5, and 5" without providing a clear rationale.
9. Check the consistency of data units throughout the text
10. Figure 7 shows a high R-value (0.95), but this is for training. Please provide the R-value and RMSE for the Testing/Validation set specifically to demonstrate the model’s generalization capability.
11. The classification of construction waste types from GF-2 imagery is presented without validation.
12. In the current version, the text, labels, and legends within Figures 1-11 are too small and unclear. The content of the images is not clearly visible. The authors should replace these with high-resolution images where all axes, legends, and map elements are clearly legible, allowing the reader to follow the analysis without difficulty.
13. There are no references from 2024 or 2025 in the current reference list. The most recent references are from 2023 (Refs. 4, 5, 23, 25, 34, 38, 42, 44, 56). This indicates that the literature review may not be fully up to date, and including more recent publications from 2024 and 2025 would strengthen the manuscript's currency and relevance. In addition, several references are missing page numbers.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript investigates the estimation and spatial distribution of decoration waste in Henan Province by integrating GF-2 high-resolution remote sensing imagery with a Back Propagation (BP) neural network model. The study aims to provide a technical framework for urban waste management and policy-making by predicting generation trends through 2030 and identifying high-accumulation zones via ArcGIS. A Major Revision is recommended to address the following technical and methodological issues.

*   Providing further details in Section 2 regarding the specific variables designated as inputs and outputs, as well as a clear definition of the model construction and a data processing workflow, would significantly improve the structural clarity of the methodology (Section 2, Lines 115–171).
*   The authors should include a comprehensive description of the programming environment (e.g., Python 3.x, MATLAB) and the specific version of libraries or toolboxes (e.g., Scikit-learn, PyTorch, or MATLAB Neural Network Toolbox) used to implement the machine learning processing (Section 2.3, Lines 144–157).
*   Clarifying the exact sources of the raw data and providing a statement on data availability or a more detailed description of the acquisition process is necessary to ensure the study is reproducible (Section 2.1, Lines 116–124).
*   It would be beneficial to explicitly detail the mathematical integration or summation method used to derive the cumulative projection of 49,827,200 tons by 2030 from the annual predicted values (Abstract, Lines 26–27).
*   Specifying the exact acquisition dates for the GF-2 remote sensing images and the resolution parameters for the multispectral and panchromatic bands would clarify the precision of the fusion process (Section 2.2, Lines 127–131).
*   The manuscript would be strengthened by a more technical description of the "machine learning recognition" algorithm used for image extraction—such as whether Random Forest, SVM, or a specific CNN architecture was employed—along with the size of the training and validation samples (Section 2.2, Lines 134–137).
*   The authors should provide a theoretical justification for using a BP neural network with four hidden layers for only four input variables, as such a deep configuration for a small input set may increase the risk of overfitting (Section 2.3, Lines 155–157).
*   Confirming the specific software version and training parameters (e.g., momentum, epoch limits, or transfer functions) beyond the learning rate would allow other researchers to verify the results of the "newff" function (Section 3.2, Line 200).
*   Including a sensitivity analysis on the 0.1 t/m² generation coefficient would help quantify how fluctuations in this primary assumption impact the final predicted volume of decoration waste (Section 2.4, Lines 168–169).
*   Establishing a standardized set of criteria for the "manual visual interpretation" phase would clarify how the authors mitigated human bias when distinguishing decoration waste from other types of construction muck or demolition debris (Section 2.2, Lines 135–136).
*   The discussion regarding the 2020 error fluctuation in Figure 8 could be expanded to explain how the model was adjusted to ensure that this pandemic-related anomaly did not skew the long-term trend projections toward 2030 (Section 3.3.1, Lines 228–232).
*   Addressing the discrepancy between the "10%" general proportion mentioned in the abstract and the specific findings for Pingdingshan (6.29%) and Xuchang (8.24%) would improve the internal consistency of the results (Section 3.1, Lines 179–182).
*   It is suggested that the authors provide the specific formula used for data normalization (e.g., Min-Max or Z-score) to ensure the neural network’s input environment can be accurately replicated (Section 3.2, Lines 200–204).
*   Clarifying the statistical method used to define the five generation levels in Figure 10—such as Jenks Natural Breaks or Equal Intervals—would add necessary rigor to the spatial distribution analysis (Section 3.3.3, Lines 254–257).
*   Implementing a robust cross-validation procedure, such as k-fold validation, would demonstrate that the reported R=0.95463 goodness of fit is consistent across different data subsets and not an artifact of a single training-test split (Section 3.2, Lines 215–217).
*   The authors might consider updating the literature review to include more recent studies from 2024 to align with the journal's current scope and to further emphasize the "Sustainability" implications of their findings in the context of the circular economy.
*   Improving the legibility and scale bars of the maps in Figure 10 would enhance the reader's ability to interpret the regional characteristics of the high-production areas mentioned in the text.
*   Refining the terminology throughout the manuscript to consistently distinguish between "decoration waste" and general "garbage" or "construction waste" would prevent confusion regarding the specific waste stream under investigation.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

All comments have been addressed.

Reviewer 4 Report

Comments and Suggestions for Authors

The reviewer thanks the authors' the effort taken on the review.
The manuscript is recommended for publication.