Application of ATH9010 in Monitoring Rice Diseases

author: Callum

2024-09-18

Application of ATH9010 in Monitoring Rice Diseases

Rice is one of major food crops and plays a crucial role in the survival and development of people. The growth status of rice directly affects the yield and quality of rice, making the health of rice a major concern for millions of farmers. Among the various factors impacting rice growth, diseases are the most significant disturbance. The invasion of various diseases can reduce rice yield or cause widespread lodging overnight, making early monitoring and scientific control of rice diseases essential.

This study utilizes a UAV to acquire hyperspectral data for monitoring rice diseases through machine learning and deep convolutional neural networks.

Rice Field

First, hyperspectral images are preprocessed. Then, deep neural networks are used to accurately extract rice fields. Following this, spectral information and disease characteristics of the rice are analyzed. Finally, PNN (Probabilistic Neural Network) is employed to precisely monitor disease levels, providing valuable references for rice management.

Flowchart

Rice disease monitoring is categorized into four levels: healthy, minor disease, moderate disease, and severe disease, based on the severity of the disease affecting the rice.

Physical Diagram of Rice Disease Levels

The process includes steps from the raw image to leaf segmentation and then to disease spot segmentation.

Lesion Segmentation

Accurate estimation of white leaf spot proportions is fundamental for assessing disease severity. Using disease severity to characterize rice bacterial leaf blight not only reduces prediction errors of spot proportions in models but also helps in developing an intelligent breeding decision mechanism by integrating disease onset times. Disease severity levels for the validation and test sets were determined based on PNN model predictions of spot proportions. The accuracy of disease severity prediction for the validation set remains higher than that for the test set, at 90.90%, with a micro F1 score of 0.90 further demonstrating the stable predictive capability of the PNN. The confusion matrix indicates that the image segmentation model tends to output results with a higher spot proportion than the actual. The accuracy for determining disease severity in the test set is 85.00%, with most misclassifications occurring in lower disease severity levels.