Comment 1. The dataset was collected in Jiangsu Province. Considering that the changes in geographic context may affect the model's performance, how can you guarantee that the model trained with this dataset can be generalized to other regions?

Response: We compared the samples from Gaofen-2 and Beijing-2 images, and found that PV panels exhibit similar characteristic in high-resolution imagery and the main difference comes from the backgrounds. The changes in geographic context will inevitably affect the performance of deep networks. To mitigate such effects, we tried to collect PV samples covering as many backgrounds as possible to enhance the generalization ability of deep networks trained by our dataset. The installed PVs in Jiangsu province are distributed on various land covers, such as, sparse shrubs, low-density grasslands, reservoirs, ponds, saline alkali lands and rooftops, making our dataset representative for most cases. Besides, some skills in deep learning community can be adopted to guarantee the generalization ability in other regions, such as transferring learning, cross-domain feature representation. In completely different areas where direct application may fail, using a small number of samples in the target area to fine-tune the pre-trained model can also achieve satisfactory results. Therefore, although the samples were collected in Jiangsu Province, there is no need to worry too much. Furthermore, our work will continue, and we plan to collect more PV samples in China and around the world through automatic deep learning algorithm and manual discrimination.

Comment 2. The dataset is composed of Gaofen-2 and Beijing-2 imagery, which include a near-infrared band in addition to RGB bands. Does this mean that the model trained with this dataset can only be applied to the satellite imagery with a specific combination of spectrums? Are these two imagery representative of the current common satellite imagery?

Response: Actually, all samples in our dataset are composed of red, green and blue (RGB) bands of used satellite and aerial images. Our experiments (Section 4.1) also demonstrate that RGB bands are enough for deep networks to distinguish PV panels from various backgrounds. Therefore, the near-infrared band of Gaofen-2 and Beijing-2 images is not contained in our samples. We have made it clear in Section 3 as “The shapefile of polygonal annotations was converted to a raster that has the same spatial resolution as satellite or aerial images. The raster and original red, green and blue (RGB) images were then seamlessly cropped into tiles at a fixed size by referring to the sampling grids.” The comparison of PV samples from Gaofen-2 and Beijing-2 imagery shows that different PV panels exhibit similar characteristic after histogram equalization. Thus, we believe that the two satellite imagery are representative for common applications.

Comments:

1 – According to the classification system introduced in Table 1, the PV dataset introduced in this study may including different categories of PVs. However, in the result part, all the categories were classified as a whole. The authors may want to show the influence of different resolutions on classifying different categories of PVs, but the “classification system” used here may confuse. My suggestion is to change the term “classification system” in the title of Table 1.

Response: Thank you for your suggestion. In order to support more applications and give full play to the data value, we divide all samples into different sub-categories according to their background land use or roof type. In the result part, we focus on validating the impact of different resolutions on PV classifications; hence, all sub-categories are classified as a whole. To avoid confusion, the term “classification system” is replaced by “organizational structure” in the revised manuscript, that is, the revised title of Table 1 is “Organizational structure of our PV dataset.”

2 – The names of segmentation networks in the manuscript should be checked. For example, “DeepLab v3+” was taken as “DeepLab v3” in Line. 189, Page. 11.

Response: Thank you very much. We have carefully checked and corrected the names of the three segmentation networks in the manuscript. “DeepLab v3” should be “DeepLab v3+” and the revised sentence is “On average, DeepLab v3+ achieved an IoU of 0.873 for flat concrete PVs and 0.927 for steel tile PVs.”

3 – This study gave suggestions on the selection of image resolutions for the classification of different PVs. The differences in the classification results of images with different resolutions may be related to the size of the target features and input samples, because the semantic segmentation networks are generally sensitive to the size, shape, and receptive fields. It may be interesting to give a quantified result on selecting image resolutions and input sample sizes for target features with different sizes in the future study.

Response: Thank you for your suggestions. In this study, we recommend to use PV08 for concentrated PV, PV03 for distributed ground PV, and PV01 for distributed rooftop PV so as to achieve the best segmentation results. However, the quantitative result on selecting image resolutions and input sample sizes for PVs with different sizes might require samples from more satellite and aerial images with different resolutions. Therefore, this interesting suggestion is only discussed in Section Conclusion of the revised manuscript. The related part is “Besides, this dataset may contribute to a diversity of other research and applications related to PV. For example, the segmentation networks are generally sensitive to the observational size and shape in the receptive field; hence, it is valuable to quantitatively explore the general guidelines on selecting image resolutions and input sample sizes for PVs with different sizes.”

4 – Authors used images with different resolutions to extract PVs in the study area. However, it may be interesting to see the classification results when combining multiscale features from images with different resolutions. The fusion of multiscale features can be further discussed in the future.

Response: Thank you for your suggestions. This paper presents two simplest applications based on our PV dataset. Now that multi-resolution dataset has been established, more complex cases are feasible. We hope that in the future our dataset will support more valuable and interesting research, for example, investigating whether a network can be established to fuse multiscale features from images with different resolutions to achieve synchronous identification or segmentation of multi-scale PVs. We have added this point into the Conclusion in the manuscript as “Whether a network can be established to combine images with different resolutions to achieve synchronous identification or segmentation of multi-scale PVs is also of great interest.”

Comments:

1) For the segmentation experiments (Section 4.1), the samples are divided into 80% training set and 20% testing set. Is the division performed randomly for PV08/03/01 or separately for each subcategory? The authors should make it clear because the two approaches lead to different model performance and segmentation results.

Response: Thank you for your advice. Actually, the division is performed for each subcategory of PV08/03/01. For example, the training (testing) set for the experiment on PV08 contains 80% (20%) samples from PV08_Rooftop and 80% (20%) samples from PV08_Ground. We have made it clear in the revised manuscript as “The experiments were conducted on PV08, PV03 and PV01 dataset, respectively. For each sub-category (e.g., PV08_Rooftop, PV PV08_Ground), all samples were separated into 80% training set (from which 20% samples were used for validation) and 20% testing set.”

2) For the cross application (Section 4.2), I wonder whether authors divide training samples and testing samples in the same way as Section 4.1? The authors state that “fine-tuning means that the model was first pre-trained on PV03 (PV08) samples, then fine-tuned (fine-tuning process lasted 10 epochs) using PV08 (PV03) samples, and finally applied to PV08 (PV03) samples.” Are the samples used for fine tuning the same as those used for direct training? If yes, how will the model perform if only using a small portion samples for fine-tuning.

Response: For the cross application, we divide training set and testing set in the same way as Section 4.1. The samples used for fine tuning are not the same as those used for direct training. Only a small portion (20%) samples from the training set of the target PV dataset are selected for fine-tuning. We have made it clear in the revised manuscript as “The training set account for 80% of the whole dataset and the testing set is the remaining 20%. Only 20% samples from the training set of the target PV dataset are randomly selected for fine-tuning.”