Plants Diseases Phenotype Identification using Machines Learning

Abstract

Environmental deterioration brought on by traditional chemical weed control is a problem that affects most agricultural land. Precision farming tries to produce food sustainably by using fewer inputs and applying them accurately in both location and time. By distributing inputs precisely where they are needed, deep-learning-based classification systems, like ViT (Vision Transformer), can help to promote sustainable agriculture by detecting weeds in future agricultural systems. Unmanned aerial vehicles (UAVs) can serve as the foundation of a weed identification system. These aerial weed-detection systems are still in their infancy, so it's unknown if they'll be practical or what the relationship is between system performance and image resolution. In this study, ViT identifies weeds in UAV photographs at different resolutions. K-fold cross validation builds datasets. This study demonstrates the theoretical economic benefits of such a system for farmers and the strong correlation between resolution and performance. This was not true for all cross validation runs, pointing to a more nuanced connection between field features and performance. Previous studies found consistent results with stronger weed-detection systems. This study connects multispectral image data to model performance to show how multispectral images affect model performance. The results for the original dataset demonstrate potential for UAVbased weed management. This study can help future theory-to-practice research. Lightweight GPUs that can run real-time deep-learning models are hindering the development of a UAV weeddetection system. We employ two commercial quadrotor UAV platforms with multispectral cameras, namely RedEdge-M and Sequoia from MicaSense, to take advantage of the potential of these technologies. This dataset was curated by Autonomous Systems Lab, ETH Zurich and contains over 10,196 images of sugar beet crop taken from sugar beet fields in Eschikon, Switzerland, and Rheinbach, Germany, with a time interval of five months. Then, in order to identify the weeds, present in crop fields and take appropriate action, the models can be used by public and private organisations as well as people working in the agricultural sector