Individual tree detection from aerial RGB images using transfer learning semantic segmentation and simulated illumination template matching in the Yellow River Delta

Efficient and accurate detection of individual trees automatically in the Yellow River Delta Provincial Germplasm In-situ Conservation Area (YRD-PGICA) can acquire data such as the location, quantity, and distribution of trees to facilitate tree species protection tasks. However, the existence of a large number of vegetation with similar spectral characteristics to tree crowns, the diverse density of wild trees, and the relatively limited image data in the study area pose challenges to the individual tree detection (ITD) task. This paper proposes a workflow that combines deep learning and template matching based on high-resolution aerial images to detect and count individual trees in the YRD-PGICA. Firstly, a semantic segmentation model is trained via transfer learning to discriminate trees from other objects in the images. Secondly, tree crown templates of varied sizes are generated based on the solar illumination characteristics of the tree crowns in the images using a simulated illumination method. Finally, a constrained 2D bin packing model and a template matching method based on the structure similarity index measure (SSIM) are used to detect the trees in the segmentation results. The accuracy of the proposed method was verified on four images with different densities of trees, with F1-scores reaching 0.809, 0.805, 0.776, and 0.799, respectively, and 252, 160, 138, and 164 trees were detected. The results show that the proposed method achieves good detection accuracy and outperforms other comparison methods. The proposed workflow can be applied to ITD tasks in the YRD-PGICA based on high-resolution aerial RGB images. It also provides a method reference for ITD tasks in other areas of the Yellow River Delta.