Soybean disease identification using original field images and transfer learning with convolutional neural networks

Abstract

Meeting the growing demand for soybeans will require increased production. One approach would be to reduce yield loss from plant diseases. In the U.S., soybean diseases account for approximately 8–25% of average annual yield loss. Early and accurate detection of pathogens is key for effective disease management strategies and can help to minimize pesticide usage and thus boost overall productivity. Recent advancements in computer vision could move us towards that goal by making disease diagnostics expertise more readily accessible to every-one. To that end, we developed an automated classifier of digital images of soybean diseases, based on convolutional neural networks (CNN). For model training and validation, we acquired more than 9,500 original soybean images, representing eight distinct disease and deficiency classes: (1) healthy/asymptomatic, (2) bacterial blight, (3) Cercospora leaf blight, (4) downy mildew, (5) frogeye leaf spot, (6) soybean rust, (7) target spot, and (8) potassium deficiency. To make training more efficient we experimented with a variety of approaches to transfer learning, data engineering, and data augmentation. Our best performing model was based on the DenseNet201 architecture. After training from scratch, it achieved an overall testing accuracy of 96.8%. Experimenting with full or partial freezing of core DenseNet201 model weights did not improve performance. Neither did a deliberate effort to increase the diversity of subject backgrounds in the digital images. Models performed best when trained on datasets composed exclusively of images of soybean leaves still attached to the plant in the field; conversely, mixing in images of detached leaves on simple backgrounds reduced performance. On the other hand, data augmentation to increase representational parity across disease classes provided a substantial performance boost. Our development experience may provide useful insights for researchers considering how to best build and analyze datasets for similar applications.