Object Recognition of a GCP Design in UAS Imagery Using Deep Learning and Image Processing—Proof of Concept Study

Round 1

Reviewer 1 Report

The authors present a proof-of-concept study for object recognition of a GCP layout in UAS images by using deep learning and image processing.

The document is very well written and the authors have made an important and grateful effort to make it understandable.

The main novelty is the methodology they test based on a regular set of tools but with a clever mix of DL approach and common image processing techniques.

Although the impact of this novelty can be debated, in my opinion, it may be of interest to the scientific community. Again, in my opinion, this document can be published as is.

 

Author Response

Thank you for your comments. We have replied in a separate document.

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

As an initial comment regarding this paper, it is useful to understand how modern aerotriangulation of drone imagery is performed automatically when GCPs are utilised. In brief: 1) Universally, nowadays, drones incorporate GPS units (maybe also an IMU). 2) Camera exposure positions then provide initial (in some cases final) constraints to satisfy control requirements for aerotriangulation /block adjustment. 3) After preliminary block formation, the location in image space of GCPs can be easily computed & consequently searching within images for GCPs is also a straightforward process. 4) GCPs do not have to be physically numbered or labelled, because there need not be ambiguity because of the tight search space. 5) GCPs are then measured in image space via manual or automatic means, there being a host of different GCP designs in use from ‘checkerboard’ to circular disks, with measurements being monoscopic or multi-image (eg template matching, centroiding, ellipse fitting, etc.). 6) Final bundle adjustment can then be performed with appropriate GCP & camera station GPS weighting constraints.

Turning now to the present paper, the concept of utilising Deep Learning (DL) to recognise marked GCPs is a good and topical idea, though it remains questionable – given point 3 above – when there are productivity gains using Neural Network approaches. The DL aspects of the paper are worthy of publication, however this reviewer would have preferred to see a more comprehensive analysis of the practical merits of the approach versus traditional object (ie GCP) detection.

As regards the new GCP target design presented in the paper, this will also be of interest, as an addition to a range of other designs. He target design appears to meet requirements in the present study, though there is only a very cursory analysis of its merits in terms of image point measurement accuracy, from both manual and automated centerpoint extraction. The paper gives very little on aspects such as network geometry (how were the 11 images configured?), image scale, etc., so one cannot assess performance by the traditionally utilised accuracy criteria. Table 5 is pretty useless, without both network configuration/scale information, and without an indication of relative weighting of GCPs versus camera station.  A much more comprehensive accuracy/productive analysis is warranted here.In the absence of this the paper is too light in regard to novelty & innovation.

The most problematic component of the paper is the treatment of, and indeed the very concept of requiring image readable numbers/labels on or adjacent to the actual GCPO targets. This is totally unnecessary and is indeed a backward step for automated aerotriangulation of drone imagery. Physical labelling is problematic (certainly based on the results presented) and not needed. Maybe the adopted workflow discussed, centering around Metashape, justified the use of physical physical labelling, but there’s a certain irony here in that DL techniques are being exploited to find targets, the image locations of which can be established within initial processing stages, and yet this intelligence is not being utilised in the correspondence determination stage, so GCPs have to be physically labelled by an operator.

Recommendation: Major revision required, which pays better attention to the current state of the art in fully automated aerotriangulation with GCPs.

Author Response

Thank you for your comments. We have replied in a separate document.

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

In this study, the authors presented a proof of concept study, in which they detected GCPs in UAV images automatically using a deep learning technique. In the GCP detection, RetinaNet50 model was used. Then, the detected GCPs were used for further processing of these images.

Following are some suggestions/comments

·         Page number 2, line number 71-76 rewrite the sentences.

·         It is suggested to add a table containing the specifications of the UAS used in this study.

·         It is suggested to add flight planning parameters chosen for conducting this study.

·         Page number 4, line number 147, “we conducted six additional UAS flights…..”. It is suggested to mention flight planning parameters.

·         Page number 4, line number 148-149, “different environments, and various flight altitudes.”. It is suggested to mention details about the environments and flight altitudes.

·         Page number 4, line 150-152, rewrite the sentences.

·         In mapping/surveying applications, flying height is generally more than the flying height scenarios considered in Figure number 3. Therefore, it is suggested to choose other flying height scenarios also.

·         Page number 5, line number 166-167, training dataset is small in size, it is suggested to consider more training images which will be helpful in improving the accuracy of the model.

·         Page number 5, section 2.2.2, authors considered RetinaNet50 model for GCP detection, other object detection models can also be tried in order to check which model is performing best for GCP detection.

·         Page number 8, line number 284, typo error.

·         It is suggested to make a table of the working environment.

·         Authors are suggested to do a comparative analysis which can include the impact of manually picked GCPs and detected GCPs on the generated products (Orthomosaic, DSM, and 3D point cloud).

·         In this study 20 GCPs were considered, if 20 GCPs are manually picked it will not take much time. In order to see the impact of this type of concept (presented in this study) it is suggested to consider more GCPs.

Author Response

Thank you for your comments. We have replied in a separate document.

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The revised paper is from my point of view now OK for publication.