General comments

This is a detailed, well-written manuscript that distinctly describes the new vectorised green roofs dataset and the deep-learning method applied to automatically detect green roofs in visible aerial imagery. It is easy to gain access to the dataset and supporting code. The methods section provides useful information pertaining to hyperparameter tuning which is important for reproducibility. The results provide valuable insight into relative green roof coverage across London boroughs (e.g., Figures 4 and 5). Although this information is already contained within the datasets referenced and compared to within the manuscript (LRW2019 and London Plan AMR 16), it is evident this is the first time this data has been provided in an open access format.  

My comments are primarily minor or typographical in nature, apart from one major concern. This relates to the lack of evidence that the trained U-Net can suitably generalise to other locations or other imagery of London. One of the benefits of training a neural network is the ability to apply the tool to automatically detect the feature(s) of interest in new images.  Therefore, it is a concern if it is not possible to do this. The trained U-Net was not tested on imagery captured during different years or seasons in the year when lighting conditions may alter the appearance of the green roofs in the imagery. It is highlighted within the discussion section that the trained U-Net produced a lot of false positive results (over-predicted green roof coverage) in some Eastern boroughs of London. This is attributed to the potential use of a different collection instrument, highlighting that the trained U-Net may not be able to generalise to imagery captured using different sensors. It is suggested that the U-Net is applied to imagery captured during different years to test the model’s ability to generalise. The imagery does not nearly need to cover the whole of London but cover one or two study locations to demonstrate the ability or otherwise for the trained U-Net to detect green roofs in a variety of settings.

Below are some other minor and typographical comments:

Specific comments

Line 105- Please detail the number of images mosaiced to cover the whole London study area and the time period covering the first and last image. Also, is there information on the time of day when the images were captured?

Section 2.3- Was any pre-processing conducted e.g., pixel value normalisation?

Section 2.3- Please provide a little more information on how the trained U-Net was applied to unseen images. E.g., were images tiled or patched prior to input? If so, how were outputs mosaiced and were there any issues with predictions at the edge of the tiles?

Section 2.3- How was it defined whether a pixel was positive or negative? Was a threshold value applied to the U-Net outputs and was a consistent value applied across images/ image patches?

Line 169- was class imbalance considered within the loss functions? If so, how?

Figure A2- The orange and blue labels are the wrong way round in the legend and image description.

Figure A3- The orange and blue labels are the wrong way round in the legend and image description. There are also no green lines in the image.

Technical corrections:

Line 86- change ‘an’ to ‘a’

115- missing word between ‘hand-labelled’ and ‘are’. Should it be polygons/ datapoints?

152- ‘m’ should be italics.

Throughout- be consistent in use of ‘hyperparameter’ or ‘hyper-parameter’.

Figure 6- please add in small-scale schematic of London to show where the example images are from.

This is a well written paper that develops a deep learning-based green roofs’ mapping framework in incorporating very high-resolution remote sensing satellites. Also, this study used data augmentation to improve the accuracy of the mapping and provide a robust dataset of green roof locations and areas in London. Overall, it is helpful for others to study aerial remote sensing mapping. However, there are some problems, which must be solved before it is considered for publication. If the following problems are well-addressed, this reviewer believe that the essential contribution of this paper are important for designing sustainable buildings and studying urban microclimates.

General Comments:

In ABSTRACT: authors are suggested to refine the abstract, focusing on the novelty of the research rather than providing extensive background information. In addition, semantic segmentation models coupled with data augmentation strategies are important in this study. It is helpful to provide the validation accuracy accordingly in this section.

In INTRODUCTION: authors are suggested to start broad in the general background, then narrow in on the relevant topic that will be pursued in the paper. The introduction sections are to highlight the challenges currently faced by green roofs’ mapping research. I suggest that the first three paragraphs be summarized in one paragraph. In addition, the detailed description of the mentioned algorithms (UNet) can be moved to the second section since we are not developing new models.

In DATA and METHODS: there are many datasets mentioned and may be clearer if summarized in a table. For example, recording information such as the coverage of the study area, the date of data acquisition and the spatial resolution of the imagery. We all know that convolutional neural networks are data-driven models. Well, how many positives and negatives are there before and after using data augmentation, and is there an improvement in model performance and by how much? Also, although deep learning or semantic segmentation models are a black box, it is helpful to provide formulas. Besides, I have some doubts about the details of the algorithm framework. For example, the loss curve during optimization. Furthermore, the general area of the study area is larger than a 256*256 image patches. How did you deal with it when predicting?

In RESULTS: several tables record the model performance. Recommendations for the structure of research paper. In Tables 3 and 5, I found that the values of TP, TN, FP, and FN vary greatly. Also, despite the high accuracy of the models in Tables 4 and 6, the performance is weaker on the IOU metrics. Why is this happening?

The image segmentation algorithm used is supervised classification. As a result, the classification results are constrained by the labels, and the model's generalization is limited. Even though we are able to assign labeling tasks to individuals, we cannot classify locations where the image is occluded. Is the current weakly supervised or unsupervised segmentation a significant advancement?

Summary: In the study titled “An Open-Source Automatic Survey of Green Roofs in London using Segmentation of Aerial Imagery”, the authors use a convolutional neural network to detect green roofs over buildings in London from aerial imagery. The paper is generally well-written and the authors sufficiently describe the model architecture, hyperparameter tuning, and some of the uncertainties in the results. The dataset will be useful for future studies on the urban climate of London and for informing heat mitigation strategies. I do have a few questions and concerns about the methodology, particularly its generalizability, that should be addressed before the paper can be considered for publication.

Major Comments:

1. I am a bit confused by how the tiles were split. The authors mention that every split contains both positive and negative classes. Is there any threshold used for what fraction can be positive or negative? Or can it, in theory, be a single negative class and the rest being positive? Additionally, was only one split used for the model training, validation, and testing. This is important since different random splits of training data might produce different results. Did the authors check for consistency across random training sets?
2. Aerial imagery for summer 2019 was used for the analysis. Was this a single image or multiple images mosaiced to cover the whole area? What were the dates of acquisition of these images? How was the presence of clouds in these images accounted for? There is not enough metadata about the imagery to understand the baseline observations.
3. Related to the previous point, it is unclear how generalizable these results and the U-net are. Would one expect the results to be generally replicable using an imagery for winter? Or if we use observations from a different satellite with the same spatial resolution?

Minor comments:

1. A constraint on accuracy of the final dataset is the OS footprint data. The authors mention that it is “very accurate” in the Limitations section. It would be good to have a quantitative estimate here. Are the OS data also for 2019?
2. Table 1: Since this includes two different survey areas, would be good to also add the green roof areas as fraction of total area or total building footprint area for easy comparison.
3. Figure A1: Please add the r2 value and equation of the line of best fit here for context
4. In the discussion, the authors talk about how green roof fraction is low compared to policy proposals and available space. However, it is important to remember, and maybe something to expand upon, that green roofs are not the only roofing strategy for heat mitigation. Of note, white roofs are generally found to be more effective that green roofs for reducing temperature, and solar panels, which have also become quite popular, can generate energy for indoor cooling. These alternative strategies would be competing for that same space.