FOSS4G-Asia 2023 Seoul

Cocoa is a major global commodity, and its production involves millions of mostly smallholder farmers in developing countries. Efficient and accurate counting of cocoa fruits during harvest is essential for effective resource management, yield estimation, and overall productivity improvement. Traditional manual counting methods are time-consuming, labor-intensive, and susceptible to errors, necessitating the exploration of innovative technological solutions.

This research introduces a novel approach to automatic cocoa fruit counting using a mobile application integrated with computer vision techniques. The proposed system leverages four services: 1) acquisition of farmers monitoring data using the mobile devices, 2) develop explainable AI model for automatic cocoa fruit recognition, 3) presentation of Cocoa pod detection and counting services through a web & mobile based application and 4) collect the farmers feedback on cocoa pod count to improve the services. To do this, Mobile app for data collection was co-designed and tested with local farmers in Cameroon. This app allows farmers to take geo-tagged photographs and transfer them to the server when connectivity permits. In total, 2132 Cocoa tree photos were manually annotated and different artificial intelligence (AI) models (e.g., faster R-CNN, darknet, ResNet-50, mobilenetv3) were developed and tested for automatic counting of Cocoa pods. The Faster R-CNN model with a ResNet-50 performed better and provides accuracy above 70%. The model was deployed on a web service and provides real-time prediction through a web- & Mobile based application

The insights gleaned from this research will spawn a new generation of tools for Cocoa farmers to use. It is an efficient, non-destructive, and low-cost method which offers useful information for them to plan agricultural work and obtain economic benefits from the correct administration of resources. Nonetheless, further research is needed to address potential limitations data collection strategies (specifically variable environmental conditions, complex Backgrounds for image collection), image Annotation challenges and ensure seamless integration into the existing agricultural workflow.