Strawberry Maturity Classification from UAV and Near-Ground Imaging Using Deep Learning

Abstract

Strawberry is ranked third in the value of production of the crops in Florida, USA. Classifying strawberry maturity and monitoring strawberry growth status in the field is very critical for accurate strawberry yield prediction, efficient strawberry field management, and achieving the highest crop production. The traditional method of distinguishing strawberry maturity is based on either physical appearance or internal chemical substance content. However, the traditional method is time-consuming and costly. In this research, an automatic strawberry maturity classification system was developed for the rapid and accurate classification of different strawberry maturity stages. A state-of-the-art deep learning method, You Only Look Once (YOLOv3), which is good at small object detection, was trained and applied to detect strawberry flowers and strawberry fruit at different maturity stages. Two strawberry image acquisition methods, aerial imaging and near-ground imaging, were compared by using the same deep learning image processing method. As a result, three and seven strawberry maturity stages were classified for unmanned aerial vehicle (UAV) images and near-ground digital camera images, respectively. For UAV images, the highest mean average precision (mAP) of strawberry maturity classification was 0.88 for a test data set at 2 m, and the highest classification average precision (AP) was 0.93 for fully matured fruit. For near-ground digital camera images, the mAP of strawberry maturity classification was 0.89, and the highest classification AP was 0.94 for fully matured fruit as well. The result shows that YOLOv3 is an excellent approach for strawberry maturity classification on both image types.

Keywords

• Strawberry maturity classification
• UAV imaging
• Near-ground imaging
• Deep learning