On-tree Mature Coconut Fruit Detection based on Deep Learning using UAV images

Coconut harvesting in the Philippines is considered one of the most dangerous agricultural jobs because it is typically done by climbing the tree. Due to the height and structure of the tree, harvesting the so-called tree of life may pose fatal injuries or even death to the pickers. This paper presents an approach to leveraging Unmanned Aerial Vehicles (UAVs) to detect the mature on-tree coconut fruits. The proposed method would help set up the vision of the autonomous robots to be employed for coconut harvesting. The model used a Deep Learning algorithm, specifically the YOLOv5 Neural Network, to train, validate and test the custom dataset for coconut fruits and finally detect on-tree coconut fruits in real-time. The dataset was composed of 588 images for training, 168 images for validation, and 84 images for testing, where a DJI Mini SE drone captured all images and real-time detection scenarios. On the other hand, Python 3 Google Compute Engine backend (Tesla K80 GPU) in Google Collab was used to process the images and implement the algorithm. The investigation confirmed that the YOLOv5 model could instantaneously detect the on-tree mature coconut fruits. With an accuracy of 88.4%, the proposed approach will be of great value in eliminating the risks of harvesting coconuts in the future. The model can also be used for coconut crop yield estimation as the system mainly detects the visible mature fruits on the coconut tree. Finally, additional images with the presence of mature coconut fruits need to be collected to be used for training to improve the mAP of the proposed system.