A key deliverable of the DOTbox project was the development of a prototype specifically designed for the traceability of oil palm products within Ghana. This prototype was built using the DOTbox architecture and integrated directly with the blockchain network established earlier in the project. The oil palm traceability prototype allowed for the secure and transparent tracking of palm oil as it moved through the various stages of the supply chain, from production at local farms to processing, distribution, and, ultimately, retail. By leveraging the blockchain, the prototype ensured that all transactions and changes in ownership were immutably recorded on the ledger, providing a reliable and tamper-proof record of the product's journey. This level of transparency is particularly important in industries like palm oil, where issues such as adulteration and mislabeling can have significant health and economic impacts.
The project team designed and developed a custom Convolutional Neural Network (CNN) model, dubbed AfroPALM-Custom, specifically for detecting adulteration in red palm oil. The model was initially conceptualized to handle multi-class classification, identifying various adulterants such as Sudan dyes and Sorghum bicolour. However, recognizing the need for broader applicability. The project team pivoted to a binary classification approach, focusing on distinguishing between pure and adulterated samples.
To optimize the model's performance, the project team conducted extensive hyperparameter tuning using Bayesian optimization. This method allowed us to systematically explore the hyperparameter space and fine-tune the model's architecture and training parameters for optimal results. The final AfroPALM-Custom model was trained with the best configuration identified, focusing on minimizing classification error and maximizing the model's ability to detect adulteration accurately.
Reported to the FAR-LeaF programme by Eric Tchao
