BP神经网络的盐沼湿地无人机LiDAR点云强度校正  

Correction of drone LiDAR point cloud intensity for salt marsh wetlands based on BP neural network

作　　者：张曙 刘昆波 谭凯 刘帅 朱卫东[3] 魏小岛 ZHANG Shu;LIU Kunbo;TAN Kai;LIU Shuai;ZHU Weidong;WEI Xiaodao(School of Artificial Intelligence,Jianghan University,Wuhan 430056,China;State Key Laboratory of Estuarine and Coastal Researsh,East China Normal University,Shanghai 20024l,China;College of Oceanography and Ecological Science,Shanghai Ocean University,Shanghai 201306,China;Shanghai Investigation,Design&Research Institute Co.,Ltd.,Shanghai,200050,China)

机构地区：[1]江汉大学人工智能学院,武汉430056 [2]华东师范大学河口海岸学国家重点实验室,上海200241 [3]上海海洋大学海洋科学与生态环境学院,上海201306 [4]上海勘测设计研究院有限公司,上海200050

出　　处：《测绘科学》2024年第10期77-86,共10页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(42171425,41901399);上海市“科技创新行动计划”自然科学基金项目(22ZR1420900);重庆市自然科学基金项目(CSTB2022NSCQ-MSX1254);中国长江三峡集团有限公司科研项目(202103552);上海勘测设计研究院有限公司科标业项目(2023HJ(83)-005)。

摘　　要：针对盐沼湿地点云强度与各影响因素间的复杂关系导致现有常规的数学物理模型难以对强度予以精确校正的问题,该文基于盐沼湿地目标反射特性、仪器光电转换原理和空间几何构造(距离和入射角)3个因素与强度数据的关联,设计了包含漫反射和镜面反射两个隐含层的BP神经网络架构,采用贝叶斯优化获取BP神经网络最佳超参数组合,建立空间几何构造与强度数据之间的映射关系,提出了一种基于BP神经网络的盐沼湿地无人机激光雷达(LiDAR)点云强度校正新方法。相比现有的改进归一化校正模型和Phong校正模型,该文方法校正后强度数据变异系数比值分别提升约4.58%和25%,校正后强度数据分类精度分别提升了6.36%和2.11%。该文方法无需进行复杂的室内和室外校正实验,也无需考虑仪器的内部光电转换机制及激光辐射传输过程,为复杂环境下多平台和多回波LiDAR强度数据的精确建模及盐沼湿地地物信息提取提供了数据基础。Aiming at the intricate relationship between intensity and the influencing factors complicates the accurate intensity correction using conventional mathematical and physical models,the correlation between the target's reflective characteristics,the instrumental photoelectric conversion principles and the spatial geometry(distance and incidence angle)of the salt marsh wetland was leveraged in this paper,then a novel method of the drone light detection and ranging(LiDAR)intensity correction for salt marsh wetlands was proposd by employing a back propagation(BP)neural network architecture with two hidden layers representing diffuse and specular reflection.In the proposed method,Bayesian optimization was employed to identify the optimal hyperparameters for the BP neural network,thereby establishing a mapping relationship between spatial geometry and intensity data.The results demonstrated a correction accuracy improvement of approximately 4.58%and 25%over the existing improved normalized correction model and Phong correction model,respectively.Furthermore,the classification accuracy by the corrected intensity data of the proposed method were increased by 6.36%and 2.11%compared to those by the two exisiting methodoriginal intensity data.Notably,this method obviated the need for complex indoor and outdoor calibration experiments and does not require consideration of the internal photoelectric conversion mechanism or the laser radiation transmission process.Consequently,this technique offered a reliable data foundation for the precise modeling of multi-platform and multi-echo LiDAR intensity data,facilitating the extraction of feature information from salt marsh wetlands in complex environments.

关 键 词：无人机LiDAR 盐沼湿地 回波强度 强度校正 神经网络 

分 类 号：TN958.98[电子电信—信号与信息处理] P231.23[电子电信—信息与通信工程]