Urban individual tree crown detection research using multispectral image dimensionality reduction with deep learning

Abstract

The dimensionality reduction processing of multispectral data is of considerable importance to deep learning-based single-tree crown detection research. However, how to use the appropriate dimensionality reduction method to improve the accuracy of single-tree detection is rarely discussed. In this work, an unmanned aerial vehicle equipped with a multispectral camera was used for aerial photography to collect multispectral images of ginkgo tree species in the research area. The original multispectral images were used to generate five different data sets through feature band selection, feature extraction, and band combination method for training three classical deep learning networks: FPN-Faster-R-CNN, YOLOv3, and Faster R-CNN. Based on the characteristics of the band selection method, the red, green, and near-infrared bands combined with different types of target detection in the network have the best results. The FPN-Faster-R-CNN network detection accuracy is up to 88.4% for ginkgo trees. The blue, red, and near-infrared band combination obtained by the OIF index has the highest amount of information but the lowest average network accuracy at 79.3%. Experimental results show the following. In the different dimensionality reduction methods, if the color and background of the target object in the image after dimensionality reduction are obviously different, and the contour is clear, the deep learning network can obtain better results in tree crown detection. However, the information content of the image itself has a limited effect on the ability of the deep learning network to detect tree crowns. In this study, the dimensionality reduction method of multispectral images is analyzed, providing an important experimental reference for the deep learning-based single-tree crown detection.