This repository is only to show the working of the plant disease detection project. All the development codes are kept confidential.
This project in collaboration with University of Florida is developed for quick, cheap and reliable on-site disease detection in plants. Since lab testing is not feasible in remote locations, this type of application is essential as a first stage diagnosis.
User simply attaches a detachable microscopic(30x magnification) lens to their smartphone and captures images of the leaves/fruits of the plant. Prior research conducted at the macroscopic level has resulted in high validation metrics, but low detection capabilities in the field.
17 biotic and abiotic diseases are taken into consideration of tomato plant because of their high economic importance.
Multiple custom CNN architecture models are trained for lower latency and weight and work in a chained fashion to determine the final output. Models can predict whether the provided image is of the upper or lower part of the leaf (Important for tomatoes since diseases are more prominent on the upper part). Models can also reject the image if it is blurry.
- 94% testing accuracy with 98% F1-score was achieved on test data obtained from fields in Florida. Further field validation is currently going on to expand the demographics.
- Received grants worth $25,000 from Microsoft AI for Earth.
WebApp is built using Streamlit. If identified, outputs are prediction confidence for each class received from the model output. A feedback mode is provided designed for pathologists to complete the data flywheel and the images are stored based on the received feedback in an Azure Blob Storage.