Rice Seedling Detection in UAV Images Using Transfer Learning and Machine Learning
, Ming-Der Yang 1,2,3,*, R. Saminathan 3, Yu-Chun Hsu 1,2, Chin-Ying Yang 2,4 and Dong-Hong Wu 5
Round 1
Reviewer 1 Report
The author addressed all of my comments appropriately.
Author Response
Thanks for confirmation!
Reviewer 2 Report
The proposed research focus on detecting rice seedlings in paddy field using several machine learning approaches.
This study has used significant amount of UAV image dataset for training models that detect a tiny size of rice seedlings.
To evaluate the proposed models' applicability, they used many images taken under various imaging conditions.
However, in the accuracy evaluation, data(Aug 20, 2019) showing particularly low accuracy were reported as a result due to the size of the target object. This data requires a precise analysis of why the accuracy is low, and improvement is needed accordingly.
Author Response
Thanks for the reviewing points.
1. The description of the reason of low prediction accuracy with the data acquired in Aug. 20, 2019 is added to line 475-480.
The data of Aug. 12 and Aug. 20, 2019 were acquired on the 17th day and 25th day since seedling transplanting, respectively. Also, the field was fertilized on the 19th day since seedling transplanting. According to the rice growth calendar [66], the stage of rice seedlings on Aug. 20, 2019 was in the middle active-tillering stage, during which the rice seedlings are growing rapidly with more canopy covers. Therefore, the size of rice seedlings from the nadir perspective is obviously different from the training data so to reduce the precision, recall, and F-1 score.
2. The methods to improve the prediction were addressed in line 528-532.
The further study will focus on detecting rice seedlings with more various image conditions, such as illumination, tone, color temperature, blur, and noise. The models can be retrained using these additional images to adapt to more image changes. Also, optimizing the model parameters to reduce computational time and increase prediction accuracy are needed to enable models to be deployed in environments with tight resources.
