Plant Species Classification and Biodiversity Estimation from UAV Images with Deep Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Find attached

Comments for author File: Comments.pdf

Comments on the Quality of English Language

N/A

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

ABSTRACT

row 1: "ecosystem", not "Ecosystem"

row 3: "RGB images"? no infrared bands?

row 4: "autumn", not "Autumn"

row 6: "biodiversity", not "Biodiversity"

row 13: "F1 Score"? define here this index, please


KEYWORDS

in alphabetical order, please


INTRODUCTION

rows 34-36: the Authors state that "Unmanned Aerial Vehicles are well suited for continuous monitoring studies as they make surveys much faster and increase the surveyable areas by reaching into zones that are hazardous or difficult to access". But what about satellite data (Ikonos, Quickbird, Sentinel etc.)? Why UAVs and not instead satellites that are much cheaper and also provide spectral info in the infrared? I know that the spatial
resolution of UAVs can be much higher, but the lack of infrared bands and the costs in terms of money and time are critical

rows 38-40: the Authors state "However, UAV surveys produce a large quantity of data that, with typical spatial resolutions in the range of cms per pixel over tens of Hectares, makes even human analysis of UAV-acquired images impractical". This drawback does not occurr with satellite data

rows 69-73: the Authors state  "We took advantage of the natural changes in vegetation coloration to make the problem of plant species classification less challenging: as the data collection in this study was done exclusively in autumn, the plants are easier to classify". So the proposed methodology is applicable only in autumn, is this correct? In effect, without infrared data it is very difficult to properly classify vegetation

rows 73-90: the Authors state "We now present the names (Japanese, scientific and English) of the nine species considered in this study". Unclear, does this mean that there are further species but the Authors do not consider them? If so, the resulting classification images will ignore the areas occupied by these further species and will assign them to the 9 species considered by the Authors

rows 96-105: the Authors state that "the main goals of the current study are... to evaluate the level of readiness of the technology presented for practical use, including a study on the effect that the number of training images has on validation accuracy as well as a real use case scenario". If so, why do the Authors not describe the costs of this study
(costs of UAV + image elaboration + field surveys etc.) and compare with costs and accuracies of vegetation classification by using satellite images?


- STUDY SITE

The study area is very small (20 hectares). A question: in case the study area were 100x larger, the costs for vegetation
classification with UAV images would be 100x higher? and the amount of data to be analyzed, which is already enormous, would be 100x larger?

row 129: "sites 1, 2, 3, and 5"? what about site 4?

row 130: "Fir (Abies Mariesii) covers about 87% of the area". A question: how can the Authors have this info if they
have not still classified the UAV images?


- Data collection

What is actually the spatial resolution (i.e. pixel size)
of the UAV images used in this study?

What is the size in gigabytes of this UAV dataset?

Is a supercomputer necessary to analyze such an amount of data?


- Data augmentation

Because the UAV images of this little study area already have huge size, is there
actually a need of data augmentation? How much is the size in gigabytes of the
augmented dataset?


- DISCUSSION

How much is the cost of this study for a tiny study area of 20 hectares?

Why Unmanned Aerial Vehicles images, and not satellite images (Sentinel, Quickbird, Ikonos etc.)?

Why not a comparison between results from UAV images and from satellite images for this kind of studies?

Was a supercomputer necessary to apply this methodology to a tiny study area of 20 hectares?
And to apply it to larger study areas, e.g. 10x, 20x, 100x larger areas?

Can this methology be applied only in autumn when plants are easier to classify?

Why not using infrared bands that are essential to properly classify vegetation?

Why not using spatial info from convolution filters to integrate the poor spectral RGB info?

These are key infos to evaluate the readiness of this technology (UAV + deep learning), which in my opinion is utterly overrated

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Comments about the Paper: “Plant species classification and biodiversity estimation from UAV images with Deep Learning”.

 The paper aims to automatically determine plant species by applying Deep Learning networks in RGB UAV images and then to estimate biodiversity by employed indices such as Gini-Simpson, Species Richness, etc. The results are promising regarding the image classification and the biodiversity estimation errors. However, the Methodology & Discussion sections need some improvements (see in the attached file). 

All in all, my recommendation is to accept the paper for publication, subject to minor revisions. Please find in the attached file the amendments which I believe are required prior to accepting the paper.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for replying to my remarks.

 

However I am not convinced of your arguments,

for the following reasons:

 

1) because the infrared bands and, in particular,

the near-infrared bands are absent in this study,

the poor spectral information (only 3 RGB bands, Red Green Blu)

can't be compensated by an over-resolution (1.5-2.1 cm per pixel).

Accurated classification of vegetation, and land cover classes

more in general, requires a balance between spectral and spatial

resolutions. In this study, the spatial resolution is even exaggerated,

while spectral resolution is very low (3 numbers for each pixel,

referring to the 3 spectral bands). Many satellites (whose

images can be bought at few hundreds euros) provide a much

more balanced set of data, with coarser spatial resolution 

(up to about 30 cm per pixel), but much better spectral resolution

with the infrared bands that provide essential information

for both vegetation classification and functional analyses (water content,

photosynthetic activity, carotenoids and chlorophylls content etc)

 

2) the previous point is not just an academic issue, 

instead has many effects. Monetary effect is the most

obvious. A feasible methodology for vegetation classification

requires that costs per-unit-of-area are reasonably limited,

in particular if the study area is huge. In this study,

the Authors classified a limited-in-size study area of only

20 hectares with 1.5-2.1 cm per pixel resolution. 

In vegetation studies and classifications, this

is considered a tiny study area. Woodlands and prairies usually

extend over thousands hectares, which makes the approach proposed

by the Authors economically (and, I suppose, also computationally) impracticable.

 

3) the second effect is that the Authors, due to the poor spectral

information of the UAV, can classify vegetation types only in autumn when

plants change their leaf colours due to the different contents of carotenoids

and clorophylls. This change in leaf colour can be captured also in the

visible bands using the RGB images of the UAV. But this is a serious temporal constraint to their methodology. 

In addition, I am not sure that all plant species in autumns change their leaf colour with species-specific speeds, it is possible that different plant species change their leaf colours with similar

or equal velocities, which would make classifications very difficult for other study areas. Again, the

absence of infrared bands make classifications much more challenging

 

4) the Authors state that, using a standard computer Asus Zenbook UX535LH, they can train their neural networks upon 6.5 GB of augmented data in about 16 hours. Of course I don't want in any way to put in doubt their answer, but when I apply simple neural networks (like MLP or RBF) to satellite data of 300-400 MB I have to wait 3-4 hours before my standard computer can classify them (of course, this also depends on the number of training points). With a 6.5 gigabytes dataset, I am pretty sure that my standard computer could burst. This is not a minor point, computational feasibility is another important point to judge a methodology

 

5) there's a practical application that would make UAVs better than satellite images in vegetation classification. 

Satellites view vegetation from the sky, thus they can provide only a one-layer classification of vegetation, i.e.

they can't view behind the tree canopies. Instead, UAVs could produce a multi-layer classification of vegetation, one above the tree canopy, another below the tree canopy in order to classify also herbaceous and bushy plants below the canopies. This is where UAVs could be better than satellites, but the Authors adopted a standard approach with one-layer classification (i.e. satellite-like view) and thus lose the competition with satellites.

 

Overall, the approach proposed by the Authors fails to be innovative and also fails to be practicable for large study areas.

 

A multi-layer approach to vegetation classification could help in the future the Authors to make their approach desirable     

 

Author Response

Thank you for your second revision. We believe most of the issues raised in the second revision were already answered in our first rebuttal. We understand that you were not convinced by our arguments and subsequently recommended the rejection of the paper.

As the system requires that we reply to your review also in this second revision, we want to thank you again for the time you dedicated to review our paper. The review process is necessarily limited so it is not exactly conducive of detailed discussion on research themes. Should you at any point want to explore any of the issues raised during this revision in more depth, please do not hesitate to contact us.