MBNet: Multi-Branch Network for Extraction of Rural Homesteads Based on Aerial Images
Round 1
Reviewer 1 Report
A really important application of techniques of image processing and image recognition is the recognition of ground features from remote sensing imagery. This paper examines techniques for the automated extraction of homestead outlines from UAV imagery and presents some encouraging results. The importance of these techniques to applications such as mapping and planning should not be underestimated.
This paper is clear, well-written and well-structured, presenting in detail the results from the evaluation of a new classification technique. There is good coverage of related techniques and background information. The evaluation of the results is good - it's nice to see a comparison with a good number of other techniques and also the use of homestead data extracted from a variety of landforms. A great deal of work has clearly been done! In general, I'm completely happy with the publication of this paper, but I have a few minor issues and suggestions, most easily explained as I go through the paper.
line 22 - helps the refine -> helps to refine
line 29 - division of the China's -> division of China's
I think that it's important to explain in the introduction exactly what "homesteads" are. I have a feeling that this word has many different meanings depending on your home country. In particular, you should explain the differences between homesteads and buildings. There is mention of this in the paragraph beginning on line 585, but it needs further clarification.
line 131 - Housing dense -> Dense housing
line 405 - enhancement -> modification
line 411 - Pytorch -> PyTorch
The figures are really clear and well-presented.
line 609 - suburban -> suburban images
In the conclusions, it would be good to see a mention of the applicability of this technique for mapping - I have spent many hours mapping buildings from satellite imagery for OpenStreetMap and I therefore appreciate the power of the methods that are being proposed. Also, UAV data was chosen for good reason - do you have any insight into the use of satellite data for the same purpose of homestead mapping?
Good references - comprehensive.
Author Response
Please see the attachment.
Reviewer 2 Report
The novelty of the proposed paper is highlighted by MBNet multi branch network developed by the authors used to extract rural homestead. The proposed system uses a deep convolutional neural network that extracts a detail branch, a semantic branch, and a boundary branch to enable accurate extraction of rural homestead.
The significance of the work presented within the research article is important as it enables accurate extraction of rural homestead maps that can be used for rural planning and development.
The paper follows the standard format of a research article and is well documented with related work references regarding UAV remote sensing applied to rural homestead. Within the introduction there are multiple references towards related works regarding convolutional networks and algorithms capable of extracting rural homesteads from satellite images and UAV images. The authors have proposed a multi-task learning architecture to improve detailed information and high-level semantic information. The system integrates joint loss function to generate semantic masks and boundary predictions. The Materials and methods sections provides an introduction of the study Area – Zhejiang Province. The proposed network framework is presented in Figure 2 and it integrates the three proposed branches (Boundary, Detail and Semantic) to provide accurate extraction of homestead. The Boundary Branch is well detailed within the paper, and it makes use of a point-to-point module to define the fine mask map. The results section presents the methods used to process the dataset within the experiment, the authors have cropped the images to 512x512 pixels based on a strategy that involves two steps, each cropped image was also enhanced to a certain probability during training, including random scaling, random horizontal and vertical flipping, random noise, random color dithering, etc. The experimental settings are based on Pytorch framework and the used made use of a server with 128GB RAM and 6 Tesla P100 GPUs, a total batch size of 16.
The authors have compared and analyzed the proposed MBNet algorithm against other state-of-the-art algorithms. The segmentation evaluation metrics as well as the boundary evaluation metrics achieves the best results among all other methods in mountain landforms, plain landforms and hilly landforms.
Within the discussion section the authors highlight the effects of mixed-scale spatial attention module as well as the refine module of the boundary branch regarding the performance of the proposed network. The boundary branch improves the semantic segmentation results of other branches, this enables accurate identification of individual homesteads when these are one next to another. The Point To Point Module effectiveness is also presented within the discussion section.
The conclusions section is based on the findings and results obtained by the authors for the proposed MBNet algorithm that integrates various modules to increase the accuracy of extracting homesteads accurately from remote sensing images.
The proposed MBNet – Multi branch network based on deep learning to enable the extraction of rural homestead based on aerial images is presented in detail and it`s effectiveness has been demonstrated by comparing it with other state-of-the-art models.
The proposed work has a high interest for readers and scientists that are working on image extraction algorithms and models that integrate deep learning techniques.
The authors present the effectiveness of the proposed Multi Branch Network to define an accurate and automatic rural homestead management system.
Author Response
Thank you very much for your recognition of our work! We will continue to work hard in this field.
Reviewer 3 Report
The newest cited reference is published in 2021. We are in the middle of 2022, tens of papers in the same research area were published in 2022. Please update the bibliography.
Some references are not complete such as:
Ghanea M, Moallem P, Momeni M. Building extraction from high-resolution satellite images in urban areas: recent methods 773 and strategies against significant challenges. Taylor & Francis, Inc.
There is no date.
Please check the reference list.
Abstract:
Please remove the following section from the abstract
“Accurate extraction of rural homestead maps is of great significance for understanding 8 and planning rural development. In recent years, significant progress has been made in the application of deep learning technology in the field of building extraction from high-resolution remote 10 sensing images. However, the current mainstream algorithms are not satisfactory in feature extraction and classification of homesteads, especially in complex rural scenarios.”
Because the abstract does not need an introduction. Please focus in the abstract on explaining the suggested approach and the result accuracy estimation.
Please do attention to the capital letters “Multi-Scale Attention”. They are not necessary for this phrase. Please check all the text.
Please don’t use we, our, us. Please check all the text.
Introduction
The commonly used statistics of rural homestead-related information mainly depend 33 on the traditional methods such as field surveys and surveying and mapping. Please cite a reference.
Since these 34 methods suffer from a large workload, long cycle, but low efficiency, it is difficult to meet the country's needs for homestead statistics and management. Please cite a reference.
Line 55: Please replace the word “pioneering”
Fully Convolutional Networks (FCN) "is" the pioneering work of deep learning in the field of semantic segmentation. Add reference.
Please underline the novelty and motivations of the suggested approach.
- Materials and Methods
2.1. Study Area
Please add a transition paragraph between two consecutive titles. Please check all the text.
Please release the phrase “this proposed technology” by “the suggested approach”.
You said:” some representative geomorphological area 127 images are selected as the study area (Figure 1).”. Unfortunately, this figure is vague, please cancel it or replace it with a representative and clear one.
Where is section number 2.2? Please check the section numbers.
Methodology
You said:” proposes a novel convolutional neural network, MBNet”. Why do you suggest a new approach though they are a heap of approaches available in the literature?
Figure 2 is vague and incomprehensible, please explain in the caption the significance of all its elements and variable. This figure is essential, it needs a lot of details and explanation. the same note is applicable to Figure 3.
Please add references for all equations taken from the literature.
Line 394: how do you suggest the threshold value.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
