UAV-based sustainable orchard management: Deep learning for apple detection and yield estimation

This article presents a method for automated apple counting using high-resolution images obtained from unmanned aerial vehicles (UAVs). The YOLO11 architecture, specifically models from YOLO11n to YOLO11x, was employed for fruit detection. Key steps included creating orthophotos, segmenting data into tiles, training a convolutional neural network (CNN) with transfer learning and data augmentation, and merging results. Images were captured using a DJI Mavic 3 Multispectral drone with a 20 MP RGB camera. Data augmentation including flipping, hue adjustment, blurring, and Tile 8×8 transformation increased the dataset from 11 to 2,000 images with 51,797 objects (34,383 apples and 17,414 fallen apples). The YOLO11x model achieved the highest performance metrics: mAP@50 = 0.816, mAP@50-95 = 0.547, Precision = 0.852, and Recall = 0.766, demonstrating its effectiveness in complex, high-density orchards. The YOLO11n model, with lower computational demands, is suitable for resource-limited environments. The method maintains geospatial alignment and visualizes fruit distribution across the orchard. An experimentally determined correction coefficient will account for fruits hidden from the camera, enhancing the accuracy of yield estimation. A Tkinter interface displays detection results and summary data for each orchard section. Future work includes integrating multispectral data and 3D modeling to enhance precision. These findings highlight the potential of deep learning to automate orchard monitoring and yield assessment.