Inspiration
Many Ghanaian farmers face major losses due to crop diseases, but most lack access to early diagnosis tools or agricultural experts. Internet access is also limited in many farming regions. We were inspired to build an AI-powered mobile solution that works completely offline, enabling farmers to detect and classify diseases in key crops like cashew, cassava, maize, and tomato using just their smartphones.
What it does
AgricScan is a mobile app that uses an on-device AI model to detect and classify crop diseases from images taken by a phone camera. The app works entirely offline, giving farmers real-time results even in remote areas without internet access. It supports 22 disease classes across four major Ghanaian crops.
How we built it
We trained a MobileNetV3 Large model using the CCMT dataset, which contains over 24,000 raw images and 100,000 augmented images of crop diseases. During training, we further enhanced augmentation at runtime by applying random rotations and cropping to increase robustness. After achieving good accuracy, we quantized the model and converted it to TensorFlow Lite for mobile deployment. The mobile app was built to capture images, run inference locally, and return results instantly.
Challenges we ran into
- Handling class imbalance and uneven data quality
- Keeping the model small and fast enough for low-end Android devices
- Ensuring accurate predictions under varying lighting and image angles
- Fully optimizing the pipeline to work offline without external dependencies
Accomplishments that we're proud of
- Built an end-to-end AI solution that runs entirely on a mobile device
- Achieved high accuracy on crop disease detection with a lightweight model
- Successfully integrated model inference into a working offline mobile app
- Made a practical tool tailored to real-world challenges in Ghanaian farming
What we learned
- How to fine-tune and optimize MobileNetV3 for mobile use
- The importance of runtime data augmentation to simulate real farm conditions
- Efficient model quantization and deployment with TensorFlow Lite
- Building user-friendly mobile interfaces that integrate with AI models
What's next for AgricScan
- Expand support for more crops and disease types
- Add multilingual support and voice instructions for accessibility
- Integrate GPS tagging for disease tracking and early warning systems
- Collaborate with agricultural agencies to test and distribute the app to farmers in rural Ghana
Log in or sign up for Devpost to join the conversation.