Bamboo Plant Classification Using Deep Transfer Learning with a Majority Multiclass Voting Algorithm
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript presents an approach to bamboo species classification using the majority multiclass voting algorithm. A variety of machine learning methods are investigated to search the optimal solution to the problem. Particularly, deep learning models and the ensemble of them is proven to outperform others based on the emprical experimental results. Details of my comments/suggestions are as follows.
-- The language needs to be improved throughout the manuscript. E.g., there are "subjects" missing for several sentences in the abstract.
-- The ensemble learning is simple yet effective in the baboo classification problem, however, combining the prediction results of multiple deep models could be computation intensive. Alternatively, one can use different data augmentation techniques or hyper-parameters to diversify the same model for better ensembling effect [1].
[1] A Baseline for Multi-Label Image Classification Using Ensemble Deep Convolutional Neural Networks
-- It seems DS1 has the best performance with machine learning algorithms. It is necessary to discuss the pros and cons of different datasets for the same problem.
-- It is counterintuitive that GoogleNet performs much worse than AlexNet as shown in Table 9. It is strongly suggested to give more details of model training for different deep models (e.g., learning rate, batch size, epochs, etc.).
-- Why not choose ResNet50 or ResNet100 in the study?
Comments on the Quality of English Language
The English language needs substantial improvement.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have submitted an interesting manuscript dealing with the image based classification of bamboos.
General comment: The topic of the manuscript is interesting and contribution is clear. However, the proposed method is not too novel. Namely, it was adapted from other computer vision tasks. The authors should mention that pretrained convolutional neural networks, their fine tuning, and combining their results/voting are common methods in the image processing literature, such as visual quality (No-Reference Image Quality Assessment with Convolutional Neural Networks and Decision Fusion, 2022), mammogram classification (Transfer learning from chest X-ray pre-trained convolutional neural network for learning mammogram data, 2018), or brain tumor classification (Brain tumor classification in MRI image using convolutional neural network, 2020).
Other comments: i) In Table 2, the authors give several characteristics of the database. Besides the table, plots about the distribution of height, diameter, etc. ii) Since deep learning involves a lot of experiments, the publication of training curves would be nice. iii) Figure 10 is not ideal, it cannot be seen which are the higher or the lower performances. Could you devise another plot to visualize the results? iv) In Table 6, the authors outline the main characteristics of the used pretrained CNNs. It can be seen that the image input size is 227x227 or 224x224. Did the authors resize the images in the database according to this? It was not clear to me from the manuscript.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors answered my questions. The manuscript is well written and contains a large amount of experimental results. I recommend this manuscript for publication.