Vision Transformer-Based Early Detection of Leaf Diseases in Crop Plants Using Multispectral Imaging

The study introduces a Vision Transformer
(ViT)-based solution of which the first step is to detect
early onset of any disease on the leaf of crop plants
with the help of multispectral imaging, which will
further improve precision agriculture. The framework
combines sophisticated spectral feature extraction and
self-attention to acquire spatial and spectral
dependencies among different bands. Multispectral
crop leaf data provided by experimental analysis was
found to have an accuracy of 97.86, precision of 96.42,
and recall of 97.11, which was higher compared to the
traditional deep learning architecture of CNN and
DenseNet. The usage of Near-Infrared (NIR) and Red
Edge bands was very instrumental in detecting the
onset of disease symptoms in different environmental
conditions. It was seen that the proposed ViT
architecture was highly computationally efficient as
shown by the 0.37 seconds per-image inference time
and is therefore a viable ViT architecture that could
be deployed in the fields in real-time. In general, this
framework will guarantee proper, timely, and
automated monitoring of crop health, as well as
enabling reasonable decision-making in sustainable
agricultural management.