Remote Sensing and Data Mining Techniques for Assessing the Urban Fabric Vulnerability to Heat Waves and UHI

Comment: Reviewer 1-Why 2003 but not 2015 heatwave?- The  built-up  parameters  referring to the building  typologies and compactness exploited in the photointerpretation of training urban fabric areas included in the transect, are based on various urbanistic characterization and planning activities including field data acquired mainly during the first years of 2000, especially for more dynamic periphery areas, as reported in the cited bibliography [32, 33]. So in order to be more compliant with field information the 2003 heatwave situation was selected. In addition in 2003 there are many evidences about the heatwave impacts on Rome area.         -How do you define the transect - test area?- The transect area of  about 20X5 was previously selected based on previous knowledge about the range of representative built up typologies present there,  given this  preliminary requisite the term “test area” was introduced in the photo-interpretation phase without any other meaning referring to subsequent classification/clustering steps.  -Introduce references to various spectral (vegetation) indices.-Two additional references  [20, 21] were introduced according to the reviewer observation and with subsequent changes at lines 115-116.  -Why do you use PCA for dimensionality reduction and how do you cope with cross-correlation of the 70 or 74 variables?-Which are the 70 or 74 independent variables namely?-One of the primary goal of the PCA and feature selection introduction was to properly handle the possible cross-correlation between the 70 feature variables extracted from the three indices multilayer and panchromatic channel for supporting the object classification step.  Of these 70 varaibles, 35 are synthetically indicated in table 4: 9x three indices multilayer + panchromatic channel, -1 because the ratio isn’t applicable to this former.  Then the variable number doubles by adding the normalized version of those cited before.Four additional independent variables,  namely the three M1_L numerical distributions found during the algorithm selection (optimizing the various parameters like accuracy,  overtraining robustness and missed classes) of the object classification step, plus the altimetry were added for the subsequent models calibration phase. -Why and how do you use NDVI as a permeability proxy?-This index is mainly sensible to vegetation density but it was also widely exploited as effective indicator of sealed soil as supported by many published works [ i.e. 22, 23].  -Break down lengthy sentences and polish-up the misspelling errors.-For additional comments and suggestions, please, refer to the uploaded reviewed version of your manuscript.-According to previous suggestions extensive changes of the manuscript were carried out.   Reviewer 2Considering the dramatically augmented availability of the heat data provided by satellites, the preliminary results obtained in the present work may provide a robust contribution in urban planning. The objective of the research is not clearly stated at the last paragraph of the Introduction. The authors use the 4 last paragraphs of Introduction to describe the objective in preliminary/exploratory way and then describe the procedure they follow in extensive way. Those 4 paragraphs should be improved in a way to introduce to the problem, show the research gaps and at the end present the aim of the paper.-The introduction chapter was modified in deep according to the reviewer indications. Some indications about the research goal were improved and the paragraph more closely referring to methodology has been moved under subsequent chapter and enhanced  with a new general processing schema.Figure 1 is a copy of other publication and should be deleted but also should be mentioned and cited in the text. DoneIn Table 2 the Building typologies should be translated in English. DoneIn Table 3 it is not clear if the classes are the same as the Table 2 - Building typologies. The caption of the Table 3 was improved and an explicative sentence was added after the table definition (Figure 3 b). Some numbers appear with comma in some tables (Table 5, 7, 8,9) and in the text.            -The mistakes have been corrected.  In line 698 the paragraph is about a Figure 9 that appear some pages later and is not clearly explained. This figure should be described and discussed in more adequate way and the description should be closer to the figure.  -The position of the Figures 8 and 9 was changed to be closer to the related description, the explanation of figure 9 was improved. Many references are cited in the text without numbering. -The mistakes have been corrected,        Reviewer 3-The attention of this manuscript was focused on how the Landsat ETM+ images were classified by different classification methods such as decision tree based approaches, SVM, ANN etc. However, its link between land cover/land use classification and UHI/ urban fabric vulnerability is weak. Secondly, the manuscript is not well organized some results are mixed with materials and methods. Thirdly, I feel this manuscript overuse the term “data mining”. Basically, this manuscript needs substantial revision to make its arguments clear.-The significant correlation between the night LST distributions considered and M1_L  morpho-types built-up classes (namely C4.5S and C-SVC) of the Rome urban fabric is evidenced by the results obtained in the calibration phase by means of  stepwise regression approaches and reported in the Table 9.-The manuscript has undergone a deep restructuration according to the previous remarks;- Some additional explanations, references and descriptions of  the “data mining” term have been introduced to improve the understanding of the application in this context.