Yield estimation and smart harvesting for precision agriculture using deep learning

Abstract

Precision agriculture is one of the fastest growing fields in recent years. In this thesis, we introduce a framework that provides farmers with a yield estimation from videos of crops and provides guided assistance for harvesting across the farm by utilizing geospatial information that is collected during the recording of the crops. We perform yield estimation by using a tracking model, DeepSORT, that can keep track of detected fruits for accurate counting. We modified the original DeepSORT algorithm to work efficiently on different fruits without the need for retraining. The proposed framework also provides assistance for smart harvesting through an optimized approach for container placement across the field. Performance evaluation shows that the proposed method achieves more than 90% accuracy on a real video footage of apple trees collected by a drone from an apple orchard and approximately 94% accuracy for pumpkin counting from an aerial drone footage.