Comparative Study of Simple CNN and U-Net Architectures for NDVI-Based Rice Crop Health Assessment Using Multispectral Imagery

Abstract

This research offered a comparison of simple Convolutional Neural Network (CNN) and U-Net architectures for plant health evaluation of rice plants under pest infestation based on Normalized Difference Vegetation Index (NDVI) imagery. The paper examined the model's accuracy, efficiency, and robustness across different environmental conditions and plant growth stages. The research methodology was based on a quality audit of labeled data across three classes (damaged, slightly damaged, and healthy), mitigating data misalignment through image and mask augmentation, and standardizing model training and testing. The model was evaluated using classification and regression metrics and confusion matrices. The quality audit showed a predominance of damaged (72.8%) over other classes (slightly damaged 26.8%, healthy 0.4%), with moderate noise levels and a manageable boundary consistency. These were addressed through appropriate augmentation techniques. The model's accuracy was 0.916, Intersection over Union (IoU) 0.578, and F1-Score 0.915, with low regression errors of 0.2892 and 0.0837 for root-mean-square error and mean absolute error, respectively. The performance of the U-Net without skip connections was nearly identical, indicating that the skip connections did not significantly affect performance under homogeneous vegetation patterns. However, the simple CNN model performed worse than the other models, with an accuracy of 0.878, an IoU of 0.528, an F1 Score of 0.875, a root-mean-square error of 0.3489, and a mean absolute error of 0.1218. These results revealed the effectiveness of the U-Net model for rice health segmentation under pest-infested conditions, using NDVI images.

Plant phenotyping relevance

NDVI画像からイネの病害・健全状態を推定するCNN/U-Net手法を比較・評価しており、植物状態の取得・抽出方法が研究の中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.