A Multiscale and Multitask Deep Learning Framework for Automatic Building Extraction

Round 1

Reviewer 1 Report

(1) On Line 57, Page 2, ‘Building contour segmentation is ... Building edge extraction involves...’. it’s better to identify the tasks for contour segmentation and edge extraction clearly. Maybe building footprint extraction and edge extraction?

(2)Line 323, Page 9, ‘Currently, no open-source building dataset that supports multitask learning is available ..., this study constructed a multitype benchmark dataset RSIBE’. For one thing, it needs an additional detailed description of this dataset in the manuscript. For another, although there is no dataset for multi-task building segmentation, it’s convenient to construct such a dataset from a public dataset for building extraction, which is better to compare with other state-of-the-art methods. At last, it’s advised to construct another dataset for the comparative experiments to prove the superiority of the proposed method.

(3) Line 335, page 9. ‘YOLOv4 and Faster RCNN are only available for building detection, while Mask RCNN, YOLACT, and MultiBuildNet are capable of contour segmentation and building instance segmentation based on building detection. Why? It seems that the encoder only extracts the features, and the multitask segmentation has no relationship to the encoder. If these are not available, that seems a limitation for the multi-task framework.

(3)It’s advised to evaluate the accuracy of the extracted edge.

(4)The extraction results in Figure 6 are unclear. It’s advised to increase the resolution and mark the results significantly. Same as Figure 9.

(5)‘4.4. Experiment based on large-scale high-resolution remote sensing images’. It needs a quantitative precision comparison.

(6)The ablation experiment only compared the method including blocks a,b,c, and a+b+c. but the significance of each block is unclear， it’s advised to add the experiment result including blocks a+b,a+c, and b+c.

(7) Please fully mark the information for each author.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors of this study developed a deep learning approach to map buildings by examining the synergy between building detection, contour segmentation, and building edge extraction. This is a very interesting topic, where indeed the synergetic examination of the information derived from such approaches can contribute significantly and improve building extraction results. Kindly find attached my comments related to the manuscript.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

There is no comment on the revised version.