Early Recognition of Wheat Powdery Mildew Disease Based on Mask RCNN

Abstract

Agricultural production and growth are seriously undermined by wheat powdery mildew disease. A total of 6.75% of wheat quality has been reduced every year due to wheat powdery mildew disease. Thus, the identification of wheat powdery mildew disease is important for farmers. In this paper, a real-time system for powdery mildew disease is proposed. The real-time system uses the Mask Region based convolutional neural networks (MRCNN) model to determine the exact location of powdery mildew disease along with their infection on each wheat plant image. An entire of 6200 images has been captured from different regions of Punjab. Among all captured images the images have been preprocessed using image resizing and noise removal techniques and a total of 1500 images have been discarded due to low background and foreground. All the preprocessed images have been used for training and testing in the MRCNN algorithm to find the exact location of powdery mildew wheat disease in wheat plant. The performance of MRCNN model is measured between ground truth and predicted label image which is denoted by Mean intersections over union (MIoU).The ground truth images were labeled using a visual object tagging tool (VOTT) and these images were considered ground truth images. After training of the MRCNN model, the maximum loss is found with a minimum number of epochs. Thus, the MRCNN achieves 0.44 Mean intersections over union (MIoU) for wheat powdery disease recognition in wheat plant images. As the MRCNN achieves 96.3% classification accuracy for wheat powdery mildew disease recognition with various training losses. After the classification of powdery mildew disease, the severity in each wheat plant with percentage of infection (POI) (39.3%) has been estimated.