Digitalization and Classification of Cesare Battisti’s Atlas of 1915
Round 1
Reviewer 1 Report
The article describes the methodology of digitally acquiring thematic information from historical maps in order to use them for spatial analysis in GIS software. The main goal and methodology of the research are very interesting, but there are several issues that should be noted:
Line 22-24
I disagree with the statement that "historical maps are typically reliable in terms of precision and spatial accuracy, so they can usually be compared with modern maps not only to make qualitative assessments but also to quantify the changes occurred in time".
Explain to the reader the topic related to the non-cartometric nature of historical maps. To mention which maps we can talk about precision and spatial accuracy. How can historical and contemporary maps (geographic projection, scale) be compared? What accuracies in the context of changes over time can we talk about.
The information about the cartographic source is spread out between chapters and paragraphs
Line 36-39 - name and creator
line 83 - scale
Line 109 - no geographic reference
line 136 - pre-Rome40 datum and unprojected lat / long coordinates
The story of the creator of the map is better described (although in my opinion it is unnecessary).
Well-described methodology, especially using algorithms with citations documenting its previous use.
The results and conclusions confirming the development of the methodology for the analyzed cartographic source.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This article proposes an automatic approach for the digital acquisition of thematic information from historical maps, which has been applied to assess the historical changes in the forest coverage in Trentino, Italy. Specifically, forests history has been evaluated with the use of historical documentary sources. Data extracted by the historical sources have been compared with the current one in order to identify changes in forest density in the last century. Thus, historical cartography has become a precious information tool to reconstruct the evolution of the forest cover of a given territory.
This manuscript is well structured and the topic is very interesting. The conclusion makes sense. As for the writing, generally, the main purpose of this article is clear. However, the quality of this paper has been reduced due to the following issues:
1. Concerning on figure 3：a map legend should be appended to explain the values presented by different colors/textures/grey levels. Similar issues also occur in figure 4, figure 5, figure 6, figure 7, figure 8, figure 9. Such issues made the figures hard to be understood.
2. As a research article, the ‘methodology’ are very significant. In this manuscript, the authors have employed machine learning techniques for the classification of the images (ref. 2.5). However, based on the current manuscript (section 2.5), it is hard to find the necessary methodology details with respect to:
- what is the input/output of the machine-learning model?
- How to make the training set (training map), manually or automatically?
- How many data involved in the training set?
Given the issues identified in the manuscript, I recommend a ‘minor revision’ as my general evaluation.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The paper used OBIA techniques to extraction forests from Cesare Battisti’s
atlas.
I have several remarks to the paper:
The English has often spelling problems like Geografic X Geographic
I have a problem with the title and also the paper itself.
If the paper is about the methods and should be like methodological paper I don't see the need to do a comparison with CORINE at the end moreover what's an added-value (except hatching) against your paper: New Tools for the Classification and Filtering of Historical Maps? https://doi.org/10.3390/ijgi8100455
If the paper is about forest detection in the area of interest the title and abstract is not suitable as there is no forest detection mentioned and all introduction should be focusing on extraction of information about the forest from the historical map.
If the paper is more geographically based and the place is important there should be more info about the locality in the paper.
The DPI 72 is definitely not sufficient and I have serious doubts concerning your results. For working with scans of historical maps is standard at least 300 better 400 or 600 DPI. 72 is suitable for monitor or web.
The final comparison with CORINE land cover is not sufficient as the atlas is s 1: 500 000 and CORINE is 1: 100 000. You don't take into account the generalization.
It is available as a digital image, but it is not georeferenced and it must be classified and filtered to create a map suitable for further processing in a GIS. You already work in a GIS environment when you have an image in GRASS GIS, its better to say in further processing (without using the word GIS).
Its QGIS not Qgis
Finally, the paper is methodologically weak and need to be rearranged and rewritten. I don't see clearly the research question behind and it will be hidden also for the readers. I highly recommend rethinking the aim of the paper and to try to make it more scientifically based and focused on the particular research question.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
This paper introduced the approaches to digitize and georeference heritage maps. This topic is interesting and involves multiple technical challenges. This paper gave detailed introductions of the background, approaches, and results. The paper structure is clear and easy to follow. My major comments are related to the methods used here: 1) the proposed approaches requires a lot of hard coded parameters. Do you need to manually tune them? If yes, it will be very time consuming and make the proposed approach not general. 2) After the segmentations, how do you do the postprocessing to connect the road networks and borders? These seems to be many small segments, which may be given wrong labels or not connected to each other. 3) If the purpose of this paper is just to detect the forests, the manual drawing seems to be more efficient and accurate than the proposed approach as the map only has multiple large polygons. 4) Optical character recognition technique is a mature technique. A similar approach seems to be applied on the problems proposed in this paper. 5) deep learning-based approaches have achieved very good results on computer vision problems. The authors could have a try.
Some minor comments can be found below:
- Introduction section: The history of Cesare Battisti is not the core part of this paper. It can be shortened.
- Line 173: how does it handle colorful map?
- Line 196: it is not very clear about how the training data are organized. I felt like it is a pixel-based classification, but how is it related to the segmentation outputs? Maybe the authors can rephrase this part a little bit.
- Line 248: How will the very small areas inside letters be removed during the post-classification? After that, how will the empty area be classified and filled?
- Line 269-273: Could you give more description of how the objects are removed based on the min, mean, and max size?
- Line 267: re-classification seems not to be needed. They can directly removed.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 4 Report
Thanks for addressing my comments! It looks good to me. One minor grammar issue:
Line 204: "points class" -> "point's class"
