Improving Semantic Segmentation of Roof Segments Using Large-Scale Datasets Derived from 3D City Models and High-Resolution Aerial Imagery

Round 1

Reviewer 1 Report

The paper presents a novel approach to automated generate training datasets for building roof segmentation from aerial images. The authors proposed to extract outlines of building roof parts from existing 3D building models and overlap them with digital orthophotos for semantic labeling. The proposed method was tested by using 3D building models in south Germany and showed much efficiency in terms of data generation.

In this paper, a deep learning-based network was in further developed to proof that the generated dataset is useful.

The topic of this paper is very interesting. This kind of work is really much needed for the deep learning study, since it is of high cost now to establish a training data set for roof segmentation.

The reviewer has several comments which might be useful to improve the quality of this paper. The detailed comments are listed as follows:

1. in the introduction (literature study part), it should indicate how difficult it was to building up the training data when citing a dataset.

2. It is quite hard to understand why the paper included the part of developing a model for roof prediction. It is enough to only address the problem of automated training data generation. You can write a separate paper for the model, so that the current paper is more focused.

3. the description of how to generate outlines from 3D building models and how the overlapping on orthophoto looks like is too brief. In CityGML, a building roof contains many polygons for individual roof part. How are the polygons merged into a big polygon as the outline of the roof part? How do you consider the orientations of each polygon when generating outlines? what should we do, when there are offset (in orientation, geometries and positions) among citygml polygons and orthophoto roof parts?

4. following the previous point, it would be better to use examples to elaborate the proposed method.

5. how was the method of evaluation for the generated dataset?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This manuscript mainly presents a method of automated generating of large-scale datasets from semantic 3D city models. By using the free and open cityGML, the proposed approach can obtain a large amount of labeled data for training segmentation models with a small amount of manual labor. This work has important significance in the current era of big data. This article has a clear and coherent structure, and essentially satisfies the publication criteria. I have only a few suggestions as follows:

1. The authors could explore the possibility of releasing a partial or simplified version of the dataset without orientation information to benefit the research community.

2. The authors could compare their results with the state-of-the-art transformer-based semantic segmentation models to demonstrate their advantages and limitations. 

3. The authors should address the issue of how well the model trained on the dataset can generalize to different cities and scenarios.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript presents Improving semantic segmentation of roof segments using large-scale datasets derived from 3D city models and high-resolution aerial imagery Further development is in the terms of comprehensive knowledgebase for roof segments mapping is necessary.

The article still needs to be modified and explained.

Recommendations for addition:

First of all, I would like to congratulate the authors for their work and dedication in carrying out this study.

1.        - Table 1. need to be specify in more detail, please add some more notes for definition class name.

2.       -  2.1.4 – recommendation – add some more information about using software interface for preprocessing data-set

3.    -     Please set the overview of all using Abbreviations

Author Response

Please see the attachment.

Author Response File: Author Response.pdf