Inspiration
Plant diseases can look very similar, but the treatment for each disease can be completely different. Farmers, especially smallholder farmers, may not always have quick access to agricultural experts, which can lead to wrong diagnosis, unnecessary pesticide use, crop loss, and financial loss.
This inspired us to build AgriDoc— a system that goes beyond simply identifying a plant disease. We wanted to answer the more important question: What should the farmer do next?
What We Built
AgriDoc is an AI-powered plant disease diagnostic and prescriptive system. A farmer can take a photograph of an affected leaf, and the AI model predicts the disease from 38 crop-disease classes.
Our system uses MobileNetV2 with transfer learning to analyse leaf images. We prepared the dataset through resizing, normalization, and image augmentation before training and fine-tuning the model.
The key feature of AgriDoc is its prescription layer. Instead of stopping at the disease name, the system connects the predicted disease with appropriate treatment guidance. This creates a simple flow:
Leaf Photo → Disease Diagnosis → Cause → Treatment → Action
Our fine-tuned model achieved 95.8% validation accuracy on the current dataset.
What We Learned
Through this project, we learned how important data preparation is for an AI system. We explored image preprocessing, augmentation, transfer learning, model fine-tuning, validation, and performance evaluation.
We also learned that a high-performing AI model alone is not enough. The output needs to be understandable and useful to the actual user. This is why we focused on making the final result simple and actionable for farmers.
Challenges
One of our biggest challenges was the difference between controlled dataset images and real-world field conditions. Real farm images can have different lighting, backgrounds, angles, and disease stages.
Another challenge was the limited crop coverage of our current dataset. It does not yet include several important Indian crops such as rice, wheat, sugarcane, turmeric, cotton, and chilli.
We also had to make sure that the model learned actual disease features rather than irrelevant background patterns or image artefacts.
What's Next
Our next steps are to validate AgriDoc using real field images, expand the dataset with Indian crops, add Grad-CAM explainability, complete the treatment recommendation module, and develop a lightweight mobile deployment.
Our vision is to make plant disease diagnosis and decision support more accessible to farmers.
AgriDoc — From a leaf photo to the right treatment.
Log in or sign up for Devpost to join the conversation.