A Scalable Earth Observation Service to Map Land Cover in Geomorphological Complex Areas beyond the Dynamic World: An Application in Aosta Valley (NW Italy)
Round 1
Reviewer 1 Report
A very complete and interesting work, very useful, the only thing missing is a bit of detail in the presentation of the tables and figures.
Tables 1, 5, 7 and 10 as well as figure 3 you should improve the quality of their visualisation. Also tables 5, 7 and 4 are figures, and check the design of table 8.
Finally, please check the references, reference 61 you have to change it in order of appearance to 58, and check the style of the references, at least 40 and 52 have errors.
Author Response
Response to Reviewer 1 Comments
We would like to thank reviewers for their appropriate comments and helpful suggestions that have been carefully considered. Majority of provided suggestions highlighted gaps in the text and were really useful to improve, we hope, paper quality. In blue referees can find their comments, in red authors’ actions to reply/satisfy requests.
In particular, the synthesis of reviewers’ comments suggested a deep revision in paper organization and harmonization. Consequently, you will find some structural changes aimed at simplifying paper reading and make content more effective.
All comments were carefully evaluated and for the most of them corrections and integrations have been provided. Thank you so much for your work!
Point 1: A very complete and interesting work, very useful, the only thing missing is a bit of detail in the presentation of the tables and figures.
Tables 1, 5, 7 and 10 as well as figure 3 you should improve the quality of their visualisation. Also, tables 5, 7 and 4 are figures, and check the design of table 8.
Finally, please check the references, reference 61 you have to change it in order of appearance to 58, and check the style of the references, at least 40 and 52 have errors.
Response 1: Firstly, we would like to thanks the reviewer for his/her review and suggestion. Concerning on, tables we agree with the referee and we have improved the quality (please see into the revised manuscript).
Concerning on the references thank you we have corrected them.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Please see the attached file
Comments for author File: 
Comments.pdf
Author Response
Response to Reviewer 2 Comments
We would like to thank reviewers for their appropriate comments and helpful suggestions that have been carefully considered. Majority of provided suggestions highlighted gaps in the text and were really useful to improve, we hope, paper quality. In blue referees can find their comments, in red authors’ actions to reply/satisfy requests.
In particular, the synthesis of reviewers’ comments suggested a deep revision in paper organization and harmonization. Consequently, you will find some structural changes aimed at simplifying paper reading and make content more effective.
All comments were carefully evaluated and for the most of them corrections and integrations have been provided. Thank you so much for your work!
Point 0: In this paper, a scalable Earth Observation service to map land cover in
geomorphological complex areas are proposed. The ideas presented in this manuscript
are sound, but some problems should be carefully considered before further
proceedings.
Response 0: Firstly, we would like to thanks the reviewer for his/her kind suggestion and comments. A general comprehensive editing has been done taking into account also the suggestion proposed by the other reviewers.
Point 1: In the “Abstract”, there is too much background knowledge and conclusions,
which can be summarized and condensed, and there are few method overviews, which
could be appropriately expanded.
Response 1: The reviewer is right. We have changed the abstract as follow: “Earth Observation services able to guarantee continuous land cover mapping are becoming of great interest worldwide. Google Earth Engine’s Dynamic World represents a planetary exam-ple. This work aimed to develop a land cover mapping service in geomorphological complex areas in Aosta Valley in the NW of Italy according to the newest European EAGLE legend. As starting year 2020. Sentinel-2 data was processed in Google Earth Engine. In particular, summer yearly median composite per each band and their standard deviation with multispectral indexes were used to perform a K-Nearest Neighbors Classification. To better map some classes a Min-imum Distance classification involving NDVI and NDRE yearly filtered and regularized stacks were computed to map so agronomical classes. Furthermore, SAR Sentinel-1 SLC data were processed in SNAP to map only urban and water surfaces to improve optical classification. Also, Deep Learning and GIS updated datasets involving the urban component were adopted starting from an aerial orthophoto. GNSS ground truth data were used to define the training and the validation sets. In order to test the effectiveness of the service implemented and its methodology, the overall accuracy was compared with other approaches. A mixed hierarchical approach represents the best solution to effectively map geomorphological complex areas to overcome remote sensing limitations. In conclusion, this service may help the implementation of Europe-an and local policies concerning land cover surveys both at high spatial and temporal resolution empowering the technological transfer in alpine realities.”
Point 2: In the “Introduction”, there is a lack of introduction to state-of-the-art methods,
such as multimodal RS fusion, and deep learning classification. Here are some newer
references:
[1] Application of Google Earth Engine for Land Cover Classification in Yasuni
National Park, Ecuador.
[2] Optical and SAR images Combined Mangrove Index Based on Multi-feature Fusion.
[3] Planet-Scope, Sentinel-2, and Sentinel-1 Data Integration for Object-Based Land
Cover Classification in Google Earth Engine.
[4] Spatio-temporal-spectral collaborative learning for spatio-temporal fusion with land
cover changes.
Response 2: The reviewer is right. Therefore, to improve the manuscript all suggested papers have been included also short state-of-the-art methods (such as multimodal RS fusion, and deep learning classification). Please see the introduction and here below we report the introduction part asked: “The application of Sentinel-2, Sentinel-1 and Planetscope in Land cover mapping are rapidly spreading since 2020 [26]. Anyway, Earth Observation data application from multi-platform sensor in particular of Sentinel-2, Sentinel-1 and Planetscope are mainly focused on plain areas [27]. There is still a lack in local application mountains area. The most of classification approach are based on single one-shot classification or combined approached with two supervised classification from optical and SAR data [28]. Others are focused on multimodal remote sensing data fusion from open-access and commercial EO Data in cloud platform like Google Earth Engine, and others on deep learning classifica-tion (mainly focused in classifying a single geometrical land cover component like urban areas) [29]. Nevertheless, a mixed hierarchical approach trying to adopted different sensor and several classifications to improve the land cover quality on mountain areas are still continues to be little explored and exploited.”
Point 3: Add a paragraph at the end of the introduction with the structure of the paper. “The
rest of the paper is structured as follows. In section 2, the study area and images used
are shown ....”
Response 3: The reviewer is right. Therefore, we have included this section as follow: “The rest of the paper is structured as follows. In section 2, the study area is presented. In section 3 the materials adopted, subdivided in 3.1 Multispectral optical datasets and 3.2 Sentinel-1 SAR dataset and 3.3 GIS products and Ground data. Then, in section 4 Methods subdivided as follow: 4.1 Sentinel-2, 4.2 Sentinel-1 that include 4.2.1, Urban mask 4.2.2 Water mask, 4.2.3 Land cover legend definition and 4.3 Training set and validation set definition. Then in 4.4 section Supervised classification algorithms and 4.5 section Deep Learning using Convolutional Neural Network (CNN). Finally, in section 5. Results and discussion and section 6. Results and discussion.”
Point 4: The figures could need a couple of works. The captions are not always selfexplanatory, Table 5 and 10 seems to be named “Figures”. Sometimes, units and scales
are missing, and Fig. 2 has no legend. Also, the figures in this manuscript should be
drawn in the form of vector diagrams, such as EPS or PDF.
Response 4: We would like to inform the reviewer that all images in the zip file are in PDF and with the higher resolution with a minimum of 600 DPI. In the text we have put figures beacise we cannot include pdf image in docx anyway taking into account his/her valuable comments we have try to improve also the images into the text. We apologize if there is someone not perfect but in the zip file all are fine. In figure 2 we left legend and the table which you correctly consider figure we have retain table because of they are processing workflow.
Point 5: Lines 256-257, there are inevitably some inherent problems in SAR images, such
as shadow and target overlapping, especially in geomorphological complex areas. The
co-registration of all the data sources and geometric correction of SAR are essential to
assure the high quality of the results. This part really lacks detail.ù
Response 5: The referee is right, but this section is already present in the initial manuscript but probably it has not been carefully read. As the reviewer correctly explained geometric correction and co-registration are crucial to assure high quality especially in geomorphological complex areas. Please see table 5 and section 4.2; 4.2.1; 4.2.2.
Point 6: Table 1 should be the name of each scene product you use, not the S2 satellite band
composition. At the same time, the number of samples used for training and testing
should be declared.
Response 6: We think referee has misunderstood this part. Table 1 is the availability of bands and mask per each scene in Google Earth Engine. It is a simple description of how Sentinel-2 collection is structured in GEE. As describe in the manuscript we consider all the images ranging from 1 May 2020 to 30 September 2020 all available filtering those with clouds and shadows at a pixel level and create a composite. It is clearly expressed into the text here we report a part “A yearly median composite imagery ranging from 01-05-2020 to 30-09-2020 without clouds and shadows has been realized. The S2 data has been used also in order to create yearly harmonized filtered NDVI and NDRE stacks with a 10 days step to map woody crops.”
Moreover, in section 4.5 has been remarked the issue as follow “Regarding to WC class a supervised Minimum Distance classification (MDC) was performed including the following input datasets: yearly cloud-shadow masked NDVI stack filtered (Savitzky-Golay) [69–71] and regularized at 10 days times-steps [72] on GEE. Annual stack of the NDRE index (Normalized Difference Red Edge Index for Agriculture) following the same procedure of NDVI stack [73]”. Please see into the text.
Concerning on the number of samples for training and validation set is already available in the text. Here we report the main part of the text in which is all explained. “As previously mentioned, regions of interests (ROI) per each class were defined mostly on the field and partially by applying both a segmentation and a spectral signa-ture-photo interpretation phase. In the image below, it was depicted the distribution of the ROIs in the study area. Each ROI per class has a number of polygon upper to 250. An overall of 4300 ROIs were defined and the 70% of them were adopted as training set the 30% as validation set. In the EO local service developed, ROI detection and relative changes through time was performed by coupling a self-made semi-automatic technique.” (this part is available in section 4.3 Training set and validation set definition).
Point 7: In the overall framework Table 5, the key content of this manuscript, classifiers
4.4 in a hierarchical relationship are not well-reflected.
Response 7: The reviewer is partially right anyway in another we think the reviewer has misunderstood a part. Table 5 is a workflow explanation on how retrieve Urban SAR mask in ESA SNAP which is one of many steps described. This procedure is part of the hierarchical approach as described in table 10 and in the results anyway we have improved the hierarchical relationship as follow:
“It is worth noting that water and urban classified with SAR and Urban Deep Learn-ing were joined together with the urban and water classes mapped with the optical data to improve this classes especially in isolated mountain villages. This has permitted to im-prove these classes by performing a semi-automatic GIS procedure. During this joining phase a minimum mapping unit (mmu) of 100 m were considered. Therefore, only pixels that have this mmu was mapped as urban the other joined that do not intersect with ur-ban (both SAR/Deep Learning and optic multispectral) less than 100 m were considered as classified by the optical data.”
Point 8: In addition, to prove the effectiveness and contribution, the authors should
consider increasing the accuracy verification comparison between this workflow
processing and traditional methods or consider adding other geomorphological
complex areas.
Response 8: The referee is right. We have added a section considering a simple approach not hierarchical as suggested in order to compare overall accuracy. Please look at table 11. Moreover, its has been added an explanation as follow: “It is worth noting that to prove the real effectiveness of the suggested approach and EO services developed a comparison with traditional methods have been followed. Therefore, the overall and K-coefficient were computed per each approach.
A traditional approach that adopted only optic multispectral data were followed by performing a unique one-shot classification by using KMC. A combined approach that adopted a single KMC supervised classification with optical data considering all classes and two classification involving only urban and water with Random Forest and SAT re-spectively with SAR data. Finally, the hierarchical approach with two optical supervised classification (KMC + MD) and two SAR classification (for urban and water respectively) and Deep Learning as described in this work.”
Author Response File: Author Response.docx
