Estimating Pasture Biomass Using Sentinel-2 Imagery and Machine Learning

Round 1

Reviewer 1 Report

Thank you for the opportunity to review this manuscript. The manuscript addresses a very topical issue. This study has examined the potential of estimating pasture biomass on dairy farms using Sentinel-2 imagery. Approximately 60% of the variability in biomass was explained through integrating time series S2 images, in-situ observations and climate data in a simple deep learning algorithm. The paper is well written, scientifically sound and succinct. Based on the clarity, quality of the work, I am pleased to accept the manuscript as it is.

Author Response

Dear Reviewer,

 

Thanks very much for your careful review and positive comments. 

 

Best regards,

Yun

Reviewer 2 Report

The manuscript is on Estimating pasture biomass using Sentinel-2 imagery and deep learning. However, It is suggested the following suggestions to be considered. 1. The review is incomplete for deep learning in terms of method representation and regression is not comprehensive. The motivation is not clear, some references are suggested to be analyzed or compared the following references on biomass. Deep Learning Based Retrieval of Forest Aboveground Biomass from Combined LiDAR and Landsat 8 Data. Remote Sensing, 2019, 11(12), 1459. Estimating Forest Aboveground Biomass by Combining Optical and SAR Data: A Case Study in Genhe, Inner Mongolia, China. Sensors, 2016, 16(6), 834. Synergistic retrieval model of forest biomass using the integration of optical and microwave remote sensing. Journal of Applied Remote Sensing, 2015, 9, 096069. Assessing of Urban Vegetation Biomass in Combination with LiDAR and High-resolution Remote Sensing Images， International Journal of Remote Sensing. 2020. DOI：10.1080/01431161.2020.1820618. 2. In section 2.2.1, it is mentioned that the height of pasture is measured by instruments and then the biomass is obtained. How is this calculated? What is the specific formula or expression？ 3. For Sentinel-2 applications and Deep learning. It is suggested the authors compare with the references: Deep learning-based fusion of Landsat-8 and Sentinel-2 images for a harmonized surface reflectance product. Remote Sensing of Environment, 2019, 235, 111425. Object Detection in UAV Images via Global Density Fused Convolutional Network. Remote Sens. 2020, 12, 3140; doi:10.3390/rs12193140. A Fully Convolutional Neural Network for Automatic Building Extraction From High-Resolution Remote Sensing Images. Remote Sens. 2020, 12, 1050; doi:10.3390/rs12061050. 4. Comparative experiments on the control factors of a single DL method cannot prove the advantages of the method itself. It is suggested to supplement the comparative experiment. 5. In the analysis of the results, are comparisons with the existing literature based on the same method or based on the same application scenario? It is suggested that the author can increase the contrast of multiple methods and the advantages of the methods mentioned in the article 6. The innovation of this paper should be highlighted in introduction. 7. In this paper, the sentinel-2 multi-spectral sequential data and the deep learning algorithm are used to construct the time-series deep learning network. The benefit and evaluation of the temporal data were not seen in the results analysis. For the experimental results, please give some discussions to explain the merits and drawbacks of the proposed method. 8. In this paper, the biomass analysis of pasture land was carried out by using multiple time series data. In the experimental results, it is suggested to increase the visualized biomass spatial distribution map.

Author Response

Dear Reviewer,

Thanks very much for your thoughtful review and constructive comments. We have carefully addressed your comments one by one. Given the tight revision deadline, the major changes we have made include:

(1) Citing 13 more references in literature review and throughout the manuscript,

(2) Inserting an equation describing field biomass calculation in section 2.2.1,

(3) Highlighting the innovation of our study in Introduction section, and 

(4) Adding a new figure (Figure 9 in the revised manuscript) to further illustrate the control factor(s) of the deep learning model (model sensitivity) in this study.

 

 

 

Author Response File: Author Response.docx

Reviewer 3 Report

Dear Authors,

 

I enjoyed reading the article, it is well written  but still some modifications can be made. Please see the suggestion below.

Comments about the text:

line 195 Add properly the reference in the Reference section.

line 199 If I understood correctly, it was made downsampling of the 10m bands to 20m. Did you do a gaussian filter first prior to the decimation of the bands?
What about the aliasing effect?

line 202 The dataset was composed of images with less than 75% of cloud coverage, right?

line 263 I suggest changing Sequential Neural Network for Multi-Layer Perceptron (MLP) since it is the name of the architecture of the NN.

In machine learning, the dataset is split into training, validation, and test sets.
Early stopping is a strategy to avoid overfitting, it can be described properly in lines 286 to 293. Also, it should be made explicit that you have used MLP in a regression problem. I think it should be more precise
to readers for understanding.

line 270 change calibration for training

line 272-274 It is necessary to be explicit about "Keras" term?. These are technical terms for machine
learning, such as loss, metrics, etc. In my opinion, it should be erased.

line 322 Fix the ref error for fig 8.

line 326 Often, it is only shown test results after the proper training. If the purpose to present the difference between a model with configuration A and B, it should be presented the learning curve (loss plot) of  both training. Then mention that Experiment A or B presented the lowest loss value.

line 388 Figure 9? It is difficult to follow the section without the figure. 

line 271 Change random subset for validation set (33% of the training set).

line 277 The simple ML model.

 

General Comments:

I would like to see how was the test results with different architectures. I believe the test result could be improved with a deeper network architecture.

In the section of the sensitivity of the model related to the farms, it is clear that the model can not generalize to a new farm, thus the low R2. I believe it is worth trying to collect more data and try a bigger MLP or a different model such as SVM.

I agree with the use of multi-sensor data, perhaps integrating Landsat (offers 15m of GSD) with Sentinel-2 image could be a way to improve the availability of data.

I am concern about the R-square=0.57 on the evaluation set be categorized as a high accuracy as indicated in line 432. Can you provide more insight about on why? Can you provide the p-value of the experiments?

 

 

Author Response

Dear Reviewer,

We greatly appreciate the valuable comments in your thorough review of our original manuscript. In this revised version, we have responded your remarks in totality with serious care. Many thanks again.

 

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

The authors have properly addressed all my comments. They have added some references requested, improved the analysis by providing the missing figure 9. However, some requested comparison was left for future work and still believe that the term 'Deep learning' is not adequate since it is a shallow network used, my concerns about the value of R2 were justified.

In my opinion, the paper is now acceptable for publication when some minor changes were made, like the proper editing of tables and figures.

Author Response

Dear Reviewer, 

Many thanks for your comments. In this revised version, we have

• replaced the term ‘deep learning’ with ‘machine learning’ to accommodate your concern,
• reduced the font of all tables from 10pt to 8pt to make them smaller, and
• revised the caption of Figure 6 to “Figure 6. Indicative architecture of the Multi-Layer Perceptron Neural Network Models used in this study.”

Thanks again and best regards,

Yun Chen

Author Response File: Author Response.docx

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.