Extracting apple tree crown information from remote imagery using deep learning

摘要

Manual measurement and visual inspection is a common practice for acquiring crop data in orchards and is a labor-intensive, time-consuming, and costly task. Accurate and rapid acquisition of crop data is vital for monitoring the dynamics of tree growth and optimizing farm management. In this work, we present a technique for orchard data acquisition and analysis that uses remote imagery acquired from unmanned aerial vehicles (UAVs) combined with deep learning convolutional neural networks to automatically detect and segment individual trees and measure the crown width, perimeter, and crown projection area of apple trees. By using an UAV platform, 50 high-resolution images of apple trees were collected from an orchard during dormancy (bare branches), and then each apple tree was detected by using a Faster R-CNN object detector. Based on these results, each tree was segmented by using a U-Net deep learning network. After convex tree boundaries were extracted from the semantic segmentation results by using an efficient pruning strategy, the crown parameters were automatically calculated, and the accuracy was compared with that obtained by manual delineation. The results show that the proposed remote sensing technique can be used to detect and count apple trees with precision and recall of 91.1% and 94.1%, respectively, segment their branches with an overall accuracy of 97.1%, and estimate crown parameter with an overall accuracy exceeding 92%. We conclude that this method not only saves labor by avoiding field measurements but also allows growers to dynamically monitor the growth of orchard trees.