An Adaptive Image Segmentation Method with Automatic Selection of Optimal Scale for Extracting Cropland Parcels in Smallholder Farming Systems

Round 1

Reviewer 1 Report

General remarks:

1.    Line 131 please indicate the specific model and manufacturer of the instrument used in the test.

2.    Line 188 please explain the difference of the three bandwidths: spatial bandwidth 188 (hs), spectral bandwidth (hr), and texture bandwidth (ht).

3.    Line 388 Is the measurement accuracy inconsistent with the 3 groups of topographic variation and geometric characteristics of cropland parcels?

4.    Line 409 please explain the selection basis of the five candidate scale parameters.

5.    The discussion should include the results of other relevant literature, rather than just expounding your own views.

6.    Line 542 Has it been tested in other places except Hubei Province? If not, please do not draw relevant conclusions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

I attached the pdf file.

1.     Explain, please, why you used k-means clustering to get a coarse grid, and how did you set its parameter number of clusters. If you overestimate it, cannot you get too many small heterogeneous regions, which can influence evaluation of h_s?

2.     Fig.9, example B, under segmented cases of MR:  what is the reason that MSAOS was better than MR-145 to MR-190:  the road was clustered as heterogeneous region and was untouched during fine segmentation, or the scale automatically pre-estimated  corresponds to lower MR scale? 

3.     In  GLCM you use one occurrence shift: have you some experience with more GLCMs with different shifts- can they influence a result, or is this the only reasonable option with respect to running time? Also, k-means clustering uses optimal texture features- are they same as for MSAOS? 

4.     Fig.6 separability index: Dis and Asm also have high separability index, why are they not suitable?

5.     Fig.7  Tile 1:In row F : does black "color" mean no cropland area? Then why cyan cropland area disappeared in F (white circle), and why in the red circle a green parcel is separated into more? What is done in E and F?

6.     Can you tell, what is the overall running time of MSAOS compared to basic MS algorithm?

 

Suggestions to the text:

Because the paper is long and it contains lots of information, I suggest in 3. Methodology to add a short remark referring to 4.1. In the abstract you mention that Entropy and Homogeneity are the most important features and it may be confusing when you mention them all. 

The coarse segmentation: maybe better description of coarse and fine segmentation can be done, see Question 2.

3.2.2 Fine segmentation

In 226 you may insert short remark that you find it iteratively by increasing the radius  and at the end h_s  is evaluated from local variance of pixels clustered as homogeneous regions. (The word adaptive is often used for something that differs at each pixel, e.g. adaptive threshold, so )to my mind, it is then easier to read.

 In 235 substitute last iteration by previous iteration

 Line 470-473 is too long and not clear. I am not sure if in addition  to  Tile 8 is an appropriate English word. Also please check the numbers  for F_ab in the sentence, I did not understand how you got them.

Table 6. Some separator between PL and HIS is missing. I wondered why Tile 6 is HIS  in line 388 and then PL in Table 6.

 

 

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The presented manuscript titled „ An adaptive image segmentation method with automatic selection of optimal scale for extracting cropland parcels in smallholder farming systems“ presents an unsupervised approach to segment PlanetScope images with an adaptive parametrization, followed by a supervised classification of the resulting segments into cropland and non-cropland objects. The authors present the combination of these methods as a tool to identify individual fields in heterogeneous smallholder-dominated agricultural landscapes.

Overall the manuscript is well organized and written. The objective of the study is timely and seeks to progress the current state of the art, but major adjustments are needed before the manuscript can be published. In the following, I will elaborate on four key issues the authors need to adress before publication of the manuscript.

1) The definition of smallholder agriculture

The authors positioning their work in a smallholder-dominated agricultural setting, but a clear definition of what this consititutes and how the authors define smallholder agriculture is lacking. While the authors state that in the region “cropland parcels are small, with more than 80% of the parcels smaller than 2 ha” (line 121 – 122), it is unclear where this information comes from and how that is reflected in the field sites under study. Since wall-to-wall reference data is available (line 147-149), the authors should report detailed field size distributions per region (tile) under study.

2) Reporting accuracies

Field sizes in the complex settings (e.g. PS tile 1) visually appear highly mixed, which demonstrates that the approach can handle small and large parcels simultaneously. However, in regards to the accuracy assessment, including large parcels in the heterogeneous setting could artificially increase the accuracy scores. The authors should instead derive performance statistics by pooling the reference data, stratifying the results by field size categories and evaluate the performance in different settings of field sizes. Also, the authors should provide more details on the quantity of reference data (number of fields and field size distribution) and the methods used to generate it (which image data was used, how many interpreters, etc.), and consider adding additional metrics for a more nuanced evaluation of the segmentation results.

3) Transferability claims

While the authors claim that the method can be applied in other smallholder settings, I find this cleam widely unfounded. The authors tested their approach in small 3x3km image chips within a small region, using the same image data in each tile. Transferability of the presented approach to other regions with widely different cropping systems (land management, crop types, landscape configuration, field sizes) using imagery acquired at different dates or phenological stages deserves further research but is not at all covered here.

Moreover, the approach systematically excludes heterogeneous image patches (to my understanding of lines 216-219), thereby categorically limiting the analyses to the more homogeneous parts of the agricultural landscape, for which segmentation approaches are working more better. While this may be a reasonable approach in the region under study, this may inherently lead to the exclusion of small and fragmented smallholder plots in other regions (presence of shading trees, intercropping, etc.), for which the approach then will not work.

I advise the authors formulate the transferability claims more cautiously or remove them from the manuscript.

4) Scaling and reproducibility

The authors do not report on the computational efficiency. While the approach demonstratively works for a small region, i.e. very small subsets it, the readers are left wondering how well the approach can be scaled up to work across larger regions? Moreover, the authors must make the code for the presented method available to the readers in order to preproduce the results and test the MSAOS algorithm in different settings.

 

Please find below some more detailed comments:

-          Line 67: add “… in settings where these reference data are absent”

-          Line 136-138: Please provide more details on the pre-processing of the PlanetScope imagery in ENVI.

-          Line 239: How were thresholds a and b determined and how may these values influence transferability of the approach?

-          Section 5.1: The comparison with the MR algorithm is not needed, as this is mostly revealing the effect of determining the optimal h. Instead the authors may choose to present a comparison of their own approach with and without this procedure.

-          Line 505: Can the authors elaborate on why they did not consider multi-temporal imagery in the segmentation?

-          Line 512: In the introduction, you discarded the use of Canny edge detection due to its sensitivity to noise.  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Thank you for considering my suggestions.