How I built an AI that diagnoses plant diseases from a photo
It started with a shocking reality:
In Cameroon, there is only one agricultural extension officer for every 1,500 farmers.
That means many farmers are forced to guess when a crop shows symptoms of disease.
A tomato leaf turns yellow — is it Early Blight, Late Blight, a viral infection, or just a nutrient issue?
The symptoms can look similar, but the treatment is completely different.
That question inspired my final year B-Tech project:
PLANT Diagnosis — a cloud-based mobile platform for real-time plant disease diagnosis and treatment recommendation.
What I built
I designed a mobile system that allows a farmer to:
take a photo of a sick leaf,
get an AI diagnosis,
receive treatment advice in English and French,
and save the case for later use.
The system includes:
React Native + Expo frontend,
microservices-based Node.js backend,
Python Flask inference server for AI processing,
MongoDB Atlas databases,
and cloud deployment across Render and Hugging Face Spaces.
The AI part
I trained two separate MobileNetV2 models:
one for tomato diseases,
one for banana/plantain diseases.
Why separate models? Because the leaf structures are completely different, and combining them would reduce accuracy.
The hardest part
The biggest challenge was data quality.
I worked with 80,000+ images from Kaggle, Roboflow, PlantVillage, and Google Images, then removed duplicates and manually verified the labels.
One major discovery: many “healthy” tomato images were actually diseased.
That meant my first model was learning the wrong thing.
After cleaning and retraining, the results improved significantly.
Results
Tomato accuracy: 93.97%
Banana accuracy: 91.33%
BBTV detection: 100% precision and recall
Inference time: 10 –20 seconds
Offline support: Yes
Bilingual treatment protocols: Yes
What I learned
This project taught me that:
data quality matters more than model complexity,
real-world systems must be built for real user constraints,
and an AI product is only useful if it works in the field, not just in a lab.
Most importantly, I learned that technology can be a practical tool for supporting farmers who have limited access to expert diagnosis.
This project started as a final-year assignment, but it became something bigger:
a step toward making plant health expertise more accessible to farmers in Cameroon.
#ArtificialIntelligence #MachineLearning #AgriTech #ComputerEngineering #ReactNative #Microservices #DeepLearning #FoodSecurity #Cameroon #FinalYearProject
3 Août 2026 à 18h18
Voir sur LinkedIn →
