Accurate Measurement of Tropical Forest Canopy Heights and Aboveground Carbon Using Structure From Motion

Round 1

Reviewer 1 Report

This a is a pretty interesting work, because of the use of AUVs is not widely explored, especially for Tropical Mountain Forest. However, some issues are raised around the present research.

In general, to clarify some concepts it is recommended the use of bibliography.

Line 26: “providing the means to make rapid, low-cost surveys over hundreds of hectares, without the need for LiDAR.”

Nowadays the actual technology cheap systems are limited to fly for few minutes, limitation that is also related with the payload of the aircraft, because the inclusion of additional sensors reduce dramatically the fight time. Also with high altitudes the covered area are increased, but it is directly inverse with the pixel sixe or Ground Sample Distance – GSD) and it could depend of the purpose of the survey.

Line 99: Section Material and Methods

It is not explained the classification of individual trees with the LiDAR approach, also there is a lack of information of the generation of the tree classification using the point cloud obtained with the SFM technique. It is recommended to evaluate the results of LiDAR and UAV approach, whereas an accuracy assessment of the estimated number of trees would add value to the research presented.

Line 137: “After excluding anomalous low points (i.e. points erroneously located outside the normal point distribution)”

The outliers are considered as anomalous values, but in general it could be positive or negative values outside the normal distribution of the data, how these outliers were removed from the point cloud?

Line 152 - 156: “Topographic Position Index and Canopy Position Index”

There is no explanation of these index and their results, and it is missing the bibliography to justify it.

Line 228:

The issue corresponds to the use of Ground Control Points (GCP) and the accuracy of the product. It is supposed that UAVs that incorporate technology as Real Time Kinematic (RTK) or Post Processed Kinematic (PPK) provide high reliability, and they are expensive, but not as the LiDAR surveys. The regular UAVs has the capacity to include the position or geotags in their photos, but it still remains accuracy problems, especially in the Z axis, where the RMSE tends to be higher to 10m.

Also a problematic part is the determination of the DTM, even with the use of GCP, because it is not known real values of the terrain, as it is indicated in the present research. Whereas, the RGB camera is not an active sensor, and it does not have the capacity to penetrate the dense forest canopy and reach the real forest floor.

Line 242: “we have shown that this is primarily comprised of bias that can be almost completely removed by our corrections”

The use the linear equation could be useful for a specific region, where the ground elevation is known (DTM), but it is challenging especially in zones with fast changing topography, as it is the case for other TMF locations (different TMF regions around the world).

Author Response

Please see PDF attached

Author Response File: Author Response.pdf

Reviewer 2 Report

This article developed a model to correct error in structure for motion UAV data to more closely match LiDAR estimates of tree height with the overall objective of being able to estimate forest carbon within dense tropical rain forests.

Minor comments:

1) Do not show us models that you do not discuss in the text. If model 2 is what you are proposing as the best, just show us that. We don't need to know about every iteration of your parameterization procedure, especially if you don't see it important enough to discuss in the text.

2) You use language not fitting of a scientific journal. For example in line 329: "We were successful in reliably estimating LiDAR top-of-canopy heights accurately and with very little bias, using SFM analyses of UAV datasets." Who is to say you were successful? What is reliably? Whether you did something accurately or with little bias is subjective. My level of tolerance of bias and error are clearly very different from yours. This type of language is scattered throughout the paper.

Major comments:

1) In your introduction you state, “In tropical forests ground returns are far less frequent than those 45 from the upper canopy but are nevertheless frequent enough to measure canopy height with a typical 46 accuracy of less than 1 m [2].” First, I cannot find anywhere in your cited source to back up this claim for tropical forests. Second, I am not convinced solely by this statement that we should assume your LiDAR data for this area to be representative of the truth. You should show the difference between field data and your LiDAR and SFM data. In this way, we can know that getting closer to your LiDAR data is, in fact, improving your SFM data.

2) You should provide more statistics about how well your model matches the distribution of heights. For example, you could provide a linear error in probability space (LEPS) metric, which would tell us how well you are capturing the tails of you height distribution. For many application in forestry and forest management knowing the height of the tallest trees is even more important than the average. I do not fully understand your focus on model 1 and 2, and your disregard for your other models, when it appears the other models may work just as well or better.

Further, were other independent variables available that you didn’t/couldn’t consider, like remotely sensed spectroscopy information? For such a small area, it may be economical (subjective) to purchase 1m data to assist in your model correction.

Author Response

Please see PDF attached.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript addresses the accurate measurement of tropical forest canopy heights and above ground carbon using structure from motion (SfM). The authors apply linear regression analysis for finding the relationship between airborne laser scanned and UAV imagery constructed measurements, which they then use to correct the error contained in e.g. forest canopy height predictions (i.e. predictions made via SfM data). The research design is appropriate and the background, methods and results are clearly presented. Only minor technical problems are to be find in the manuscript which I list next: 

- Line 2: I think abbreviation SfM is more generally used. Apply this in the whole text. 

- Line 38: Airborne laser scanning is usually abbreviated as ALS, not LiDAR, even though they usually mean the same thing. If you use LiDAR in text, then maybe this term should also be clarified. Readers familiar with it can always fast forward. 

- Line 60: Change "structure from motion" -> SfM

- Figure 1: Did not find any reference to this figure in text. Should there be?

- Line 123: Maybe write "UAVS" as "UAVs"

- Figure 2: What does c and d refer to? 

- Line 134: Maybe change "Digital Terrain Models" to "DTMs"

- Line 183: Unclear what the 369 stands for.  

- Figure 3&4: Caption "SFM (SFM)"?

With these minor problems addressed I think the manuscript can be accepted for publication in the journal. 

Author Response

Please see PDF attached

Author Response File: Author Response.pdf