RGB-DSM数据融合研究——以城市樟树检测为例

RGB-DSM data fusion:a case study in urban camphor tree detection

单木检测是森林资源调查的基础工作,在森林管理与监测中发挥着重要作用。无人机RGB图像作为一种经济高效的遥感数据源,已被广泛应用于单木检测。然而,由于城市树木形状和结构的多样性以及城市森林的复杂性,基于无人机RGB图像检测城市树木效果并不理想。数字表面模型(DSM)提供地物的高程数据,融合RGB图像与DSM高程数据可有效识别城市树木,但不同融合方法对于单木检测精度的提升效果差异尚未明确。针对上述问题,本研究以城市樟树为研究对象,使用两类数据融合方法:像素级融合(IHS变换融合、Brovey变换融合和多通道组合)与特征级融合(双分支特征融合和SE-双分支特征融合),并采用Faster R-CNN模型进行单木检测实验。实验结果表明,SE-双分支特征融合的提升效果最显著,平均精度提高9.3个百分点,达91.3%。并且,在街道与绿地两类城市森林场景下,SE-双分支特征融合也表现最佳,平均精度分别达89.2%与93.9%。研究结果表明,在多树种的绿地场景下,引入高程数据可有效提高单木检测精度;在单一树种的街道场景下,高程数据提升效果有限。

Individual tree detection plays a critical role in forest inventory and monitoring.RGB images obtained from unmanned aerial vehicles(UAVs)have gained considerable traction as a cost-efficient data source for individual tree detection.Concurrently,deep learning algorithms have progressively found applications in forestry due to their remarkable effectiveness in the field of image processing.However,individual tree detection in urban areas based on UAV RGB images using deep learning algorithms remains challenging due to the diverse shapes and structures of urban trees,as well as the complexity of urban forest.Digital surface model(DSM)provides elevation data,and the fusion of RGB images with elevation data has emerged as a promising approach for identifying urban trees.Nevertheless,the improvements of different data fusion methods on the individual tree detection accuracy remains unclear.To address the challenge,this research compared the effectiveness of distinct data fusion methods in enhancing the urban camphor tree detection accuracy.The methods were categorized into two types based on distinct phase of data fusion,i.e.,the pixel-level and feature-level.Pixel-level fusion methods include intensity-hue-saturation(IHS)transform fusion,Brovey transform fusion and multi-channel fusion.Meanwhile,feature-level fusion methods encompass dual-branch feature fusion and squeeze-and-excitation(SE)-dual-branch feature fusion.Subsequently,the Faster R-CNN model and its derivatives were applied for the camphor tree detection in urban scenes.The experimental results showed that the SE-dual-branch feature fusion method notably improved the accuracy of the urban camphor tree detection by 9.3 percentage points ave-rage precision,reaching 91.3%.Furthermore,the SE-dual-branch feature fusion method achieved the highest detection accuracy in both street and green areas,reaching 89.2%and 93.9%,respectively.The research results demonstrated that the SE-dual-branch feature fusion method effectively eliminated the dimensional differences among various types of features and enhanced feature extraction,thereby significantly improving the individual tree detection accuracy.Additionally,the integration of elevation data considerably enhanced the individual tree detection accuracy in the green areas with high species diversity.Conversely,the elevation data had a limited effect in the street scene with a single tree species.This research provides a valuable reference for individual tree detection in urban areas.