A Fourier Frequency Domain Convolutional Neural Network for Remote Sensing Crop Classification Considering Global Consistency and Edge Specificity

Round 1

Reviewer 1 Report

This paper proposes a convolutional neural network called the Fourier frequency domain convolutional (FFDC) net, which transforms feature maps from the spatial domain to the frequency spectral domain. The results demonstrate that the proposed method achieves better performance on remote sensing crop classification.

Strength:

1. This paper is very well-written and easy to follow.

2. Involve the frequency spectral domain in the design of neural networks has a high novelty.

3. Ablation study is comprehensive.

However, I still have some minor questions and suggestions as follows:

1, Please provide more clear figures in Figure 12 (d).

2, In the abstract, the author claims that the proposed method could improve the robustness. Could you please add some experiments to show this improvement?

3, Please also add the computation complexity analysis of the FFDC net. such as training time and the number of parameters of FFDC net.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The paper “A Fourier frequency domain convolutional neural network for remote sensing crop classification considering global consistency and edge specificity” use a frequency-domain deep learning model to crop classification, from remote sensing information from high temporal and spatial resolution PlanetScope constellation.

 In the introduction it is not clear which is/are the objective(s) of this work.

Material and methods presents a series of shortcomings:

 

a) It is not clear how many farms (and their names) were used:

L53-54 “The study area is illustrated in Figure 1, and is comprised of six farms, Ganhe, Guli, Dongfanghong, Yili, Najitun and Dahe Bay”.

In figure 1 Dahe Bay does not appear and in its place it reads Dahewan, the same as in table 1.

L184-189: “Based on the 2022 farm planting plan provided by the local agricultural reclamation company, and a ground survey data from multiple farms in October 2022, we selected the mid-August 2022 images for Ganhe farm and manually delineated the soybean, corn and background areas as the training dataset. Data collected from Yili, Dongfanghong and five other farms were used as the testing dataset.”  Here we have one farm used for training and seven farms used for testing, quite different from the original six farms.

 In results L372-373: “The spatial distribution and frequency domain variations of the seven test regions are illustrated in Figure 7.” Here we have 7 regions (farms?) for testing, but in the figure 7 appears Ganhe farm two times (why?). Ganhe farm was presented as training dataset…. I am lost…These structure of farms appears also in Figs 8, 9 and 11

 b)  In L175-176 the authors states: “Crop classification in remote sensing is influenced by various factors, such as precipitation, soil, climate and management practices, which result in distinct morphological and textural features in the remote sensing images.” However no information about precipitation of 2022 in each farm, management practices (as sowing date, seed density, row spacing, irrigation.. etc.), soils,  was done.

 c) L190-192: “By cropping the images, we obtained a total of 2385 samples, each with dimensions of 256*256 pixels, with 1908 samples designated for the training dataset, and 477 samples for the validation dataset.” The authors do not specify how many of these cuts correspond to the different classes (soybean, corn and background or noncrop areas).

 d) The authors use a set of evaluation metrics calculated from TP, TN, FP and FN variables , which are described as: L363-366 “where True Positive (TP) and True Negative (TN) represent the number of pixels correctly predicted as positive and negative classes, respectively. False-positive (FP) and false-negative (FN) represent the number of nonobject pixels incorrectly classified as positive and the number of object pixels incorrectly classified as negative, respectively.” What is a nonobject pixels incorrectly classified as positive or object pixels incorrectly classified as negative?

 

In result L368-369 “To validate the effectiveness of FFDC, we conduct comparative analyses with Unet[27], DeepLab V3+[30], PSPnet[29] and MACNnet[13]”. These models were not presented in materials and methods. Also the models Unet, DeepLab V3+ and PSPnet were designed and tested for Biomedical Image Segmentation(Unet) and Image Segmentation (DeepLab V3+) and Scene parsing (PSPnet); the two last  tested on PASCAL VOC 2012  and Cityscapes benchmarks, a collection of RGB images from digital cameras that have different kind of animals and vehicles or cityscapes among others. Only MACNnET was designed and evaluated for crop classification from remote sensing images.   L369-370 “In this paper, we randomly select seven regions from six farms to evaluate the model's performance.” From this sentence I am completely lost and I cannot continue the evaluation of the draft.

Minor comments are highlighted in the draft.

 

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear Authors,

you have presented a very interesting paper on the use of convolutional neural network in the Fourier frequency domain for remote sensing evaluation of plants. The use of modern digital farming practices is very important in practical agriculture, as it allows for better crop management and accurate understanding of plant mechanisms. Please respond to the following comments/questions, the consideration of which will help improve your work.

1 Is your proposed method universal? can it be used to work with different crop species? In any research area? does it require re-validation?

2. how many cases were used to build and how many to validate the models?

3. please provide basic data on the soil environment of the study sites. What are the sites?

4. Admittedly, you have prepared a separate part of the work, calling it a discussion, but he recommends refining it. Please cite at least 10 publications from the last 10 years dealing with similar topics. Try to compare in more detail the results of your own research results with those obtained by other authors.

5. Please add DOI references in the references (where possible).

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors solved all my concerns and I recommend accepting the paper.