Abstract:
Estimation of yield in olive production is very important in order to enhance the overall
management and economic returns from olive orchards. Conventional methods of yield
estimation are time consuming and most of the times they are prone to mistakes. This thesis
presents a new method that applies image segmentation and artificial intelligence on the
edge devices to enhance the estimation of the olive yield. This work suggests that with the
assistance of adopting the U-Net architecture, it is conceivable to segment olives from
photographs taken in-field, under natural light. The images were segmented manually to
obtain a special dataset with binary masks that allow to clearly distinguishing olives from
the rest of the background and, consequently, provide accurate predictions of olive yield.
These images were used to train the U-Net model, and evaluated on dice score and the
mean average precision (mAP). This thesis also presents the benefit of image segmentation
over the object detection approach in managing the issues of natural orchard environment
and the importance of data processing on the edge devices such as real-time data analysis
and minimal data transfer. From the findings of this study, it was concluded that the
proposed method has the potential of greatly improving the precision of the olive yield
estimation while at the same time being a more efficient approach that could be used on a
large scale in the management of the orchard. The trained algorithm was deployed on
NVIDIA Jetson Nano and tested on field images. Positive results along with low latency
makes this research a vital push in the government's already planned increase in local olive
production. The research belongs to the agricultural technology domain, as it identifies
how AI can be applied in practice in actual farms.
