Crop Disease Detection

We're tackling food insecurity with AI. Our solution is a deep learning-powered image classifier that identifies 22 crop diseases across 4 major food crops and subsequently generalized for all crops.

Comment

Inspiration

Agriculture in Ghana—and across Africa—is experiencing an unprecedented stagnation in innovation and improvement. Yet, with the rising demand for agricultural products, there is a pressing need for equal or even greater innovation within the sector to maximize yields and minimize losses. One promising solution is to harness the power of Artificial Intelligence (AI) and Machine Learning (ML) to address critical challenges, one step at a time. Today, we focus on leveraging AI to detect crop diseases on farms, a key step toward smarter and more resilient agricultural practices.

What it does

Our product essentially classifies disease classes in crops from the leaf of the crop

How we built it

We began by training models to obtain the necessary weights for disease classification. To boost accuracy, we built separate models using Enhanced MobileNet for each crop type. These models were integrated into a custom-built Model ContextProtocol (MCP), which acts as an intermediary between user queries and the AI tools. The MCP includes components like a leaf detector for isolating leaves and a classification tool for identifying diseases. Working alongside an AI agent, the MCP streamlines responses based on the system architecture and the nature of each query.

Challenges we ran into

  1. Training Limitations Due to Resource Constraints Long training times and frequent failures were caused by limited computational resources. As a result, we had to reduce the number of training epochs, which directly impacted the final model accuracy and performance.

  2. Large Model Sizes and Deployment Issues The enhanced MobileNet models, though accurate, were relatively large, making deployment difficult—especially on low-resource edge devices or mobile platforms commonly used in rural farming communities.

3.Energy Consumption and Cost Running large AI models can be energy-intensive and costly, especially on servers or devices without hardware acceleration like GPUs or TPUs.

Accomplishments that we're proud of

  1. We are proud to have built a functional AI system, successfully integrated with both the user interface and backend, and practically tested for real-world use. This our first AI real word application

What we learned

  1. Machine learning can be difficult 😂
  2. Model Control Protocol
  3. End user Design

What's next for Crop Disease Detection

Natural Language Processing : Locan language to English and English to Local language

Built With

Share this project:

Updates