Augmenting Rice Plant Health Detection using Transfer Learning and Multispectral Imaging

Rice is a highly valued crop worldwide as it provides food for 80% of the population. Identifying rice plant disease is essential for improving agricultural losses in the long run and enhancing food security. The study suggests a suitable method for assessing the health of rice plants using multispectral imaging and transfer learning. Through multispectral imaging it captures a wide range of spectroscopic wavelengths that are beyond the human eye. That makes it easy to detect subtle changes in plant physiology which could indicate a disease. The proposed technique involves capturing multispectral images of rice plants grown under limited stress. Transfer learning is used to build discriminative features that distinguish between healthy and diseased rice plants using pretrained convolution networks (CNNs) using multispectral data. In this work, a U-Net method for rice plant disease identification is proposed and compared with other popular classification techniques such as SegNet and LinkNet. The results show that the U-Net model has a high segmentation accuracy and can capture fine grained characteristic accurately. The proposed model has IoU score 89.1%, Dice Coefficient value is 91.3% and accuracy is 94.2%.