Cost-Minimizing System Design for Surveillance of Large, Inaccessible Agricultural Areas Using Drones of Limited Range

Round 1

Reviewer 1 Report

This is a very well written manuscript, easy to follow, and well structured. In this paper, the authors provide an algorithmic solution that permits to identify the position of the charging stations (CS) and the trajectory of the drones that need to surveil particular areas. This solution is applied to inaccessible agricultural areas, which makes the paper suitable for this special issue. The authors have run simulations and conducted test performance.

1- The introduced techniques are interesting from the application perspective but there is generally a lack of novelty in the provided solutions. The authors use very classical and well-known ideas and simplify the problem by decomposing it into:
a. find CS location;
b. find the best tessellation;
c. find path.

2- The reviewer was hoping for a coupling perspective of the above components. In particular, solutions that allow to solve the location and tessellation, or location and path, or all three together. As it is now, it is an oversimplified set-up, for an interesting problem.

3- In the paper, the authors are using the favorable locations for the charging stations. What would happen if the favorable locations were not provided? Can the provided solution still optimize over this? Can we calculate which are the locations of the favorable solutions? This would really improve the provided solution.

4- The authors should provide more details on how the Integer Linear Programming was solved. For now, it is unclear, and this is known to be NP-hard. The way it is solved would make a big impact on the overall provided solution and the cost-minimization that the solution states to provide.

5- Finally, the review would like to see a Pareto Frontier with the number of stations vs time. This would permit to understand the optimal solution provided per number of stations. Once this frontier is provided then the policymakers (or Argo-company owner) can decide what suits best for them.

6- There is a general lack of appropriate related work reference regarding the optimization part. The authors could start by checking the following papers:
[1] Optimal sensor placement and motion coordination for target tracking. S MartíNez, F Bullo. Automatica 42 (4), 661-668.
[2] Sparsity-promoting optimal wide-area control of power networks. F Dörfler, MR Jovanović, M Chertkov, F Bullo. IEEE Transactions on Power Systems 29 (5), 2281-2291.
[3] On cooperative patrolling: Optimal trajectories, complexity analysis, and approximation algorithms. F Pasqualetti, A Franchi, F Bullo. IEEE Transactions on Robotics 28 (3), 592-606.

Overall, the problem is an interesting one, but generally, this paper presents very weak contributions due to the lack of novel approaches used in the solution and the over-simplified problem set-up.

Author Response

We repeat the reviewer comments, with our responses in italics.

1- The introduced techniques are interesting from the application perspective but there is generally a lack of novelty in the provided solutions. The authors use very classical and well-known ideas and simplify the problem by decomposing it into:
a. find CS location;
b. find the best tessellation;
c. find path.

“Novelty” does not necessarily lead to better performance. It is not uncommon in the technical literature that sophisticated solutions are used for the sake of novelty, without benchmarking against more classical approaches that sometimes turn out to give better results. As it stands, as far as we are aware our solution to this problem is the best that can be found in the literature. We have also cited several recent references who use similar approaches for related problems. Note that although our approach uses a classical framework, we have made several performance-enhancing design innovations that are not at all obvious (see our response to the next point).

2- The reviewer was hoping for a coupling perspective of the above components. In particular, solutions that allow to solve the location and tessellation, or location and path, or all three together. As it is now, it is an oversimplified set-up, for an interesting problem.

We have added a detailed justification for our approach.  There are hard constraints on the set of  CS’s  used which must  be respected, so we deal with these first. Our choice of CS set is coupled to distance information, which serves as a proxy for mission time (we have improved the coupling from the previous version of the paper, and achieved up to 5% better time efficiency for some scenarios). Our tessellation was designed to enable a recursive algorithm that approximates the boustrophedon solution, which is known to be the best solution for infinite-range drones.

3- In the paper, the authors are using the favorable locations for the charging stations. What would happen if the favorable locations were not provided? Can the provided solution still optimize over this? Can we calculate which are the locations of the favorable solutions? This would really improve the provided solution.

We have added text explaining how our solution also applies when there  are no preferred locations. The user may supply to the algorithm a  regular or quasirandom lattice of points that fills the field of interest.

4- The authors should provide more details on how the Integer Linear Programming was solved. For now, it is unclear, and this is known to be NP-hard. The way it is solved would make a big impact on the overall provided solution and the cost-minimization that the solution states to provide.

We have added this information to the paper. We used the glpk ilp solver, which uses branch and bound. We have given a graph of run times, which shows that even in the largest instance the maximum runtime was under 100 sec. This shows that NP-hardness is not a practical obstacle in our particular application. There are many algorithms that are exponential time in the worst case, but still can be used to solve large problems. The simplex algorithm is one example. Weather prediction is also exponential time.

5- Finally, the review would like to see a Pareto Frontier with the number of stations vs time. This would permit to understand the optimal solution provided per number of stations. Once this frontier is provided then the policymakers (or Argo-company owner) can decide what suits best for them.

The Pareto Frontier has been added.

6- There is a general lack of appropriate related work reference regarding the optimization part. The authors could start by checking the following papers:
[1] Optimal sensor placement and motion coordination for target tracking. S MartíNez, F Bullo. Automatica 42 (4), 661-668.
[2] Sparsity-promoting optimal wide-area control of power networks. F Dörfler, MR Jovanović, M Chertkov, F Bullo. IEEE Transactions on Power Systems 29 (5), 2281-2291.
[3] On cooperative patrolling: Optimal trajectories, complexity analysis, and approximation algorithms. F Pasqualetti, A Franchi, F Bullo. IEEE Transactions on Robotics 28 (3), 592-606.

Several additional references to related CPP problems  have been added, including the third reference in this list. The other two references provided by the reviewer do not appear to be closely related to our problem.

Overall, the problem is an interesting one, but generally, this paper presents very weak contributions due to the lack of novel approaches used in the solution and the over-simplified problem set-up.

These points have been addressed above.

Reviewer 2 Report

Paper is written in a good form. The idea is innovative and the approach tries to solve o future problem.

Author Response

No response is needed.

Reviewer 3 Report

For starters there are so many mis-claims in the introduction that it leads me to believe that the authors have no idea what they are talking about. 

1) Farmers are using multispectral analysis to evaluate their crops, not RGB cameras.  

2) A hectare is 2.47105 Acres. 29.5 hectares are 72.9 acres and a DJI Inspire should be able to survey this area in 2 flights, not 200.  I routinely do over a 100 acres in about 18 minutes with my EBee.  While an Inspire will fly slower than the eBee it should still be capable of doing 1/2 the area in less that 22 minutes.  I've done 25+ acre plots with a Parrot Bluegrass in 1 flight with room to spare.  

3. No one is using Mavic Air drones to do multispectral analysis!

4. Might make more sense for crop spraying, but that work won't be done with DJI drones.  

Author Response

We repeat the author’s assertions, followed by our responses in italics.

For starters there are so many mis-claims in the introduction that it leads me to believe that the authors have no idea what they are talking about. 

1) Farmers are using multispectral analysis to evaluate their crops, not RGB cameras. 

We clarify that drone surveillance is not only used for crop evaluation, but for other applications as well We provide several reverences in which RGB cameras are used for agricultural surveillance. Several of the applications in references [10-29] employ RGB images. See in particular reference [30], which provides a recent (2019) comprehensive survey of agricultural surveillance applications. Please note also that our CPP does not depend on the type of sensor used, and can be applied to drones with a variety of operational parameters. 

2) A hectare is 2.47105 Acres. 29.5 hectares are 72.9 acres and a DJI Inspire should be able to survey this area in 2 flights, not 200.  I routinely do over a 100 acres in about 18 minutes with my EBee.  While an Inspire will fly slower than the eBee it should still be capable of doing 1/2 the area in less that 22 minutes.  I've done 25+ acre plots with a Parrot Bluegrass in 1 flight with room to spare.  

We have removed this example. This was a mistake on our part, and we apologize.

3. No one is using Mavic Air drones to do multispectral analysis!

In the scenario that we are using to validate our model, we are considering a Third World scenario where very low cost equipment is required, and in particular RGB imaging is used. As we have demonstrated above, many applications use RGB. We have made these points more explicit in the introduction.

4. Might make more sense for crop spraying, but that work won't be done with DJI drones

This point is addressed above. There are a wide variety of agricultural applications  requiring a wide variety of drone types, as attested by numerous references that we have provided. In the application we have chosen to use as an illustration, low-cost drones are used.

Reviewer 4 Report

• The paper is well-organized, but I believe that the authors could include more references at the introduction and discussion.
• The discussion should be enlarged in order to involve and refer to other similar studies
• Please delete the second “the” in the line 280.

Author Response

We have corrected the typo that was pointed out, and have added numerous references in the introduction to similar studies and applications.

Round 2

Reviewer 1 Report

The review acknowledges the authors' efforts to address some of the issues pointed out. Nevertheless, the reviewer believes this manuscript does not provide sufficient novel results to be published as a journal paper.

• The reviewer acknowledges the explanation about the coupling of constraints. Nevertheless, the newly provided solution/explanation is again an over-simplification of an interesting problem, which reduces the usefulness of the solution in practical and real scenarios.
• The authors did not address the issue of not provided favorable locations. The reviewer was not asking if the user can provide these favorable points, but if they can be calculated automatically without further inputs from the user.
• The Pareto frontiers provided do not provide the numbers of the needed charging station.
• The authors need to discuss their solutions with respect to the state of the art and evaluate the results by comparing it with existing solutions.

Author Response

We have repeated the reviewer's comments, with our responses in italics:

 

• The reviewer acknowledges the explanation about the coupling of constraints. Nevertheless, the newly provided solution/explanation is again an over-simplification of an interesting problem, which reduces the usefulness of the solution in practical and real scenarios.

 

We have explained in our previous response why our solution is indeed general. See also our response to the next bullet point.  

• The authors did not address the issue of not provided favorable locations. The reviewer was not asking if the user can provide these favorable points, but if they can be calculated automatically without further inputs from the user.

 

We respectfully disagree. We clearly specified what the user should do if (s)he has no preferred locations within the region. The user may simply input a dense set of potential locations. This is a standard method for discretizing continuous problems.  Given this input, then the algorithm will select from these potential locations an optimal set.

 

• The Pareto frontiers provided do not provide the numbers of the needed charging station.

The Pareto diagram gives the covered area per charging station versus mission time per area.  The number of charging stations may be calculated from the covered area per charging stations, and the area of the field. Since this calculation is trivial, we do not feel it is necessary to explain this. The reason that we presented the Pareto frontier the way we did was to show that the tradeoff is nearly independent of the field area.  If we did what the reviewer suggests, then this valuable information would be obscured.

 

• The authors need to discuss their solutions with respect to the state of the art and evaluate the results by comparing it with existing solutions.

There are no existing state of the art solutions to this problem.

Reviewer 3 Report

Revisions accepted, please publish

Author Response

The reviewer recommends publishing.

Reviewer 4 Report

The paper is improved since the last version.

I believe that the discussion should be increased to involve more references of similar studies.

The Conclusions are incomplete. There are more data to be included. The abstract presents conclusions that are not contained in the section of conclusions. The conclusions are limited in five rows and it is difficult for the reader to understand the purpose of this research.  

            Overall this revision is a slight improvement, but I feel that the paper can be improved further.

Author Response

Our responses to the reviewer's points are in italics.

 

I believe that the discussion should be increased to involve more references of similar studies.

We have included numerous references to similar studies in the introduction.  There are no existing studies that are sufficiently similar that we can directly compare our results.  So it is more appropriate to discuss other studies in the introduction, where we have put them.

The Conclusions are incomplete. There are more data to be included. The abstract presents conclusions that are not contained in the section of conclusions. The conclusions are limited in five rows and it is difficult for the reader to understand the purpose of this research.  

We respectfully disagree that the conclusions are incomplete All of the conclusions mentioned in the abstract are supported by the results. Our claim of 70-90 percent efficiency is proved by Figure 15.  There is no need to recapitulate this information in the conclusions section.

We believe that we have made the purpose of the research to be very clear. Our purpose was to design, simulate, and verify a practical, efficient system for drone surveillance of large, remote agricultural areas. As far as we are aware, this is the first system of its kind. We have employed several optimization techniques, so our solution has both practical and theoretical interest.

Round 3

Reviewer 4 Report

Accept in present form.