A machine learning algorithm to predict crown rot in organic bananas

Organic bananas are better for the environment and consumer and worker health. However, this approach makes export banana susceptible to postharvest decay during the 2- to 4-week refrigerated transport to overseas markets. Postharvest decay typically manifests itself by rot of the cutting surface of the cluster, known as crown rot. This latent infection cannot be observed before shipment. It has been observed that the occurrence/incidence of crown rot has a seasonal pattern. Conventional bananas are usually treated with a synthetic fungicide, not approved in organic agriculture. Effective organic antifungal treatments are hardly existent. The use of controlled atmosphere (CA) containers (reduced O₂/increased CO₂) instead of standard refrigerated containers seems to reduce crown rot development. However, scientific evidence is thin and CA has additional costs. We therefore studied the effect of CA on crown rot using a data set of 7000 real shipments of organic bananas. The correlation of crown rot incidence with historic weather data, the use of CA, transit time, and other parameters were assessed. The CA indeed mitigated crown rot development at higher temperature sums and longer transit times. The data were then used to develop a machine learning model that can predict the occurrence and incidence of crown rot on future shipments, permitting CA to be deployed when useful. Crown rot is expressed in two risk classes (high, low). Different algorithms were tested and compared to predict the risk class. Because most observations were in the lowest risk class, the data set was imbalanced. This caused bias and lack of prediction accuracy. Synthetic sampling was therefore used to correct the imbalance. An overall prediction accuracy of around 80% was eventually achieved. The study shows the potential of machine learning to predict the occurrence of crown rot, and thus contributes to preventing fruit decay and food waste.