Detection of crop diseases using enhanced variability imagery data and convolutional neural networks

• Crop disease identification based on short-range imaging and CNN. • The CNN is trained and used with real-life images. • Images are split into several patches resulting in a highly diverse dataset leading to a robust CNN. The timely detection of crop diseases is critical for securing crop productivity, lowering production costs, and minimizing agrochemical use. This study presents a crop disease identification method that is based on Convolutional Neural Networks (CNN) trained on images taken with consumer-grade cameras. Specifically, this study addresses the early detection of wheat yellow rust, stem rust, powdery mildew, potato late blight, and wild barley net blotch. To facilitate this, pictures were taken in situ without modifying the scene, the background, or controlling the illumination. Each image was then split into several patches, thus retaining the original spatial resolution of the image while allowing for data variability. The resulting dataset was highly diverse since the disease manifestation, imaging geometry, and illumination varied from patch to patch. This diverse dataset was used to train various CNN architectures to find the best match. The resulting classification accuracy was 95.4 ± 0.4%. These promising results lay the groundwork for autonomous early detection of plant diseases. Guidelines for implementing this approach in realistic conditions are also discussed.