RCDDF: Application Framework for Real-Time Cassava Disease Detection

Disease Diagnosis in crops is a very expensive and lengthy process, as it often involves the need for few experts to make observations, leading to delayed diagnosis, viral spread of diseases, and thus, massive economic losses. This study presents an end-to-end solution; called RCDD for early diseases diagnosis in crops via a mobile/web application, delivering diagnosis inference in real-time to application users. It explored modern state-of-the-art improvements and techniques in the Artificial Intelligence (AI) ecosystem such as transfer learning, model ensembles, TPU machines usage, TFRecords data formats for more efficient storage and training with TPUs, offline portability of models; as well as the use of modern tools such as Pytorch to develop powerful models to diagnose crop diseases. For real-time inference on applications, the study explored and leveraged recent advancements in software engineering as well as MLOps (DevOps for ML applications) to facilitate easy deployment, and updates of developed models to servers accessible via REST APIs. Given constraints with internet in remote farmland regions, RCDD also offers offline models that are compact versions of their real-time versions that work without an internet connection. This work focuses on Cassava as an example crop by collecting a large dataset of its leaf images, to develop and build deep learning models. The model achieved an accuracy of up to 94-95% in recognizing various Cassava diseases. The system’s architecture was validated through extensive experiments, demonstrating high accuracy and potential for scalable deployment across different crops. The proposed RCDDF framework provides a promising solution for reducing diagnosis costs, improving timely disease management and enhancing crop productivity in resource-constrained settings.