Estimating apple yield by detecting orchard apples in UAV acquired RGB images

By Lars Ten Kate

Abstract
Within the field of precision agriculture, yield estimation plays a crucial role in the decision-making of farmers. Traditional methods rely on manual measurements. Computer vision, particularly artificial neural networks, can help automate fruit counting and yield estimation. These newer methods often require manual labour to acquire images which is time-consuming, inaccurate and prone to human errors. UAVs offer a solution by providing cost-effective, flexible and quick access to aerial images. This report investigates apple yield estimation by detecting orchard apples in unmanned aerial vehicle (UAV) acquired RGB images. First, an existing dataset containing UAV images is used to train a faster R-CNN model. During the training phase, the influence of reflections by the sun and shadows in the trees on apple detection are investigated by different setups of the training datasets. Lastly, a prediction by the best performing model is used in a regression to estimate apple yield. An underachieving model with an F1-score of 0.56 and an average precision of 0.45 resulted in a regression with an R2 of 0.39. Despite these low performances, the results demonstrate that a front to end set-up to estimate apple yield is feasible and deserves further investigation. As a final addition, recommendations were made to improve this front to end set-up including implementing a better annotation strategy, refining image acquisition methods, and incorporating state-of-the-art deep learning algorithms.