Few-Shot Learning for Crop Mapping from Satellite Image Time Series

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper describes a series of experiments on applying few shot learning (FSL) in crop mapping. Eight FSL methods were implemented and tested. They are Transductive Information Maximization (TIM), a-TIM, Entropy-min (transductive fine-tuning), Model-Agnostic Meta-Learning(MAML), Prototypical Networks, Baseline, SimpleShot, and MetaOptNet. The results showed that a-TIM (Veilleux et al. 2021) achieved the best accuracy among all the methods.

However, the accuracy achieved for selected crop types in scenario 1 is too low to be considered successful in classifying them in the study area, even with the best approach employed. Further analysis is needed to investigate the reasons behind this low accuracy, focusing on both data and methodological aspects.

Specific comments:

1.       Line 129 on page #3: Simpleshot à SimpleShot. This should apply to reference 31.

2.       The repository for code and instructions on setting up the benchmark dataset should have been created, even just a placeholder. There is an empty repo, https://github.com/Sina-Mohammadi/FewCrop, from the primary author’s github account. Is that the correct repo? If so, please revise the links.

3.       The classification accuracy for certain crop types, such as groundnut, maize, soybean, yam, and intercrops, is noticeably low. We need a more thorough analysis to understand the cause of this low accuracy. Simply citing the diverse agricultural landscape and high cloud cover as explanations may not be sufficient.

Author Response

See attached comments

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This article proposes a procedure for crop mapping based on supervised deep learning methods. The results are validated using three datasets obtained from France, Switzerland, and Gabon. I consider the article to be well-written, especially in the introduction and description of the method, and the graphics adequately describe the procedure. My main concern lies in the analysis of the results, particularly the quality of the values predicted by the model in relation to ground truth. Thus, in Figure 5, poor classifications are observed in the first twelve rows of the confusion matrix. I believe this should be explained in greater detail.

Additionally, I believe that the use of the Dirichlet distribution should be explained more extensively.

I believe that the confusion matrix should be explored more thoroughly.

Author Response

See attached comments

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The experimental design in this paper is reasonable, the process is clear, and the discussion is detailed. It can be accepted directly.

Author Response

We thanks the reviewer for his positive comments