A Handheld Device for Plant Disease Detection Using Multispectral Imaging

In recent years, the climate change has significantly affected the agricultural production. Maintaining the crop production is one of main concerns in agriculture. High temperature and changes of rainfall patterns enhance the spread of plant diseases. Hence it is desirable to seek for early detection of plant disease, and thus to control the spread of plant disease. Hyperspectral imaging has been proved to be an efficient tool for early detection of strawberry Anthracnose. To improve the efficiency of plant disease detection, this research aims to build a handheld multispectral imaging device for strawberry Anthracnose detection. This device uses an embedded system as the controller of the device. By placing filters in front of four miniature cameras, the images of four characteristic wavelengths are acquired. After capturing images using the handheld multispectral imaging device, images are processed to correct the effect of uneven lighting. Then by further processing the multispectral images and incorporating the RGB image of inoculated strawberry leaves, we are able to analyze the status of strawberry leaves at various infection stages. In this research, we first used the multispectral imaging device to classify the healthy and symptomatic areas in strawberry leaves. Then we further attempted to classify the status into three categories: healthy, incubation and symptomatic. SVM model was applied for classification of infection stages. For classification of healthy and symptomatic status, detection accuracy is above 90%. For classification between healthy, incubation, and symptomatic status, the accuracies are 92.2%, 68.6%, and 97.9%, respectively. The classification result of strawberry Anthracnose infection is further displayed on the handheld device as pseudo-color image so the user can easily observe the plant health condition, and so the disease management can be applied if necessary. Since the detection accuracy can be affected by lighting and shadow due to uneven surface of strawberry leaves. We propose a method to amend the effect of shadow on status classification. Through observations of the original four images and their association, a new set of images derived from the original four images was selected and tested to rectify the shadow effect. Using this new set of derived images and trained with SVM, classification accuracy for healthy status increased from 71.3% to 95.7% and the classification accuracy for symptomatic status also increased from 82.3% to 88.9%.