PP-YOLO: Deep learning based detection model to detect apple and cherry trees in orchard based on Histogram and Wavelet preprocessing techniques

Abstract

The number of technological systems such as unmanned aerial vehicles (UAVs) and artificial intelligence (AI) used in agricultural areas is increasing with the development of today’s technology. Therefore, it is predicted that spraying and fertilization processes will be carried out using AI-based drones. The implementation of these processes requires AI models. This study used the YOLOv5, YOLOv8, and YOLOv9 algorithms to detect cherry and apple trees in agricultural areas. The dataset used to build the YOLO model was generated using the DJI Mavic UAV. A data augmentation method was applied to the data to increase the number of images in the dataset. The dataset includes 2000 images of cherry trees and 1600 images of apple trees. Two approaches, Unprogressed YOLO (UP-YOLO) and Progressed and Preprocessed YOLO (PP-YOLO), were proposed in this study. UP-YOLO provides training for the YOLO models. In the proposed PP-YOLO method, the dimensions of the images are configured compared to the classical YOLO model. A spatial attention module (SAM), improves the model’s detection performance by highlighting the leaf color, leaf structure and branch texture of trees. This helps to reduce the rate of undetected objects. Additionally, PP-YOLO enhances the model’s performance by applying Histogram Equalization (HE) and Wavelet Transform (WT) image preprocessing techniques to the images. HE and WT pre-processing techniques were used to enhance the tree branch, leaf, and ground transitions and remove noise in the UAV images. While an F1 score of 94.3 % and mAP50 of 96.9 % were obtained with UP-YOLO, the YOLOv8m model with WT applied to the images in PP-YOLO obtained an F1 score of 95.8 % and mAP50 of 98.3 %. The results show that the F1 score and mAP50 of the PP-YOLO reach 1.5 % and 1.4 % higher than UP-YOLO, respectively. It was observed that preprocessing techniques increased the F1 score by 0.9 % and the SAM module by 0.6 % during the application of the proposed method. The developed deep learning model was highly accurate for cherry and apple tree detection.