Spectral-aware deep learning: enhancing plant disease detection with multispectral imaging and transformers

The conventional disease-detecting approaches tend to be based on the RGB imagery that does not capture the first unnoticeable signs and symptoms of plant infections. To address this deficiency, we propose a deep learning architecture based on multispectral imaging and Vision Transformer (ViT-B16) architecture will be proposed to perform accurate, robust, and early detection of diseases in different species of crops. It relies on the detection of six spectral bands ( RGB + NIR ) acquired by a multispectral camera and are useful in exposing pigment degradation, water stress, and other disease signs that could not be found using a conventional camera. A hand-labeled set of 4,500 multispectral images (Healthy, Early Blight, Late Blight and Powdery Mildew) is augmented significantly with data augmentation such as injection of spectral noise, rotated and brightness varied. All the images are transformed in a standard form of 224 x 224 x 6 to comply with requirements as a input. Through spectral rich image patches under a CNN model can associate patterns in space and the spectrogram domain. Also, a spectral feature fusion approach enhances the performance by linking visual indicators to unknown physiological stress indications. Transfer learning, based on ImageNet-21K pretrained weights are used to fine-tune the network, which greatly increases its accuracy, and minimizes the computational costs. Extensive tests prove the strength of the system in different field scenarios, making it possible to use it in real time, on drones, mobile scanners, and precision agriculture systems.