基于高光谱和LiDAR的黄河口湿地植被分类方法

Classification Method of Wetland Vegetation in The Yellow River Delta Based on Hyperspectral and LiDAR

摘要:

利用无人机（Unmanned Aerial Vehicle， UAV）高光谱影像（Hyper-spectral Imaging， HSI）和激光雷达（Light Detection and Ranging， LiDAR）数据开展黄河口湿地植被分类方法研究。由于高空间分辨率HSI光谱变异性强，以及LiDAR点云密度不均匀，分类结果呈现出“椒盐”现象。为了解决这些问题，本文提出了一种结合空谱特征融合和通道注意力机制的双分支卷积神经网络（SSF-C-DBCNN）。光谱注意力机制通过为每个波段分配不同的权重来减少光谱变异性的影响。空间注意力机制侧重于学习和强调特征表达能力强的密集点云区域空间信息，从而减轻LiDAR点云密度不均匀对结果的影响。最后，在双分支融合特征后引入通道注意力机制来提取更深层次的特征。利用UAV采集的HSI和LiDAR数据进行实验验证，结果表明，本文提出方法的性能优于随机森林和五种深度学习方法，分类结果更为贴合实际土地覆盖，有效地抑制了“椒盐”现象。

Abstract:

By utilizing Unmanned Aerial Vehicle (UAV) Hyper-Spectral Imaging (HSI) and Light Detection and Ranging, this study aims to investigate the classification methods of wetland vegetation in the Yellow River estuary using LiDAR data. However, due to the high spatial resolution HSI spectral variability and uneven LiDAR point cloud density, the classification results exhibit a "pepper and salt" phenomenon. To address these issues, this paper proposes a two-branch convolutional neural network (SSF-C-DBCNN) that integrates empty spectrum feature fusion and channel attention mechanism. The spectral attention mechanism mitigates the impact of spectral variability by assigning different weights to each band. Meanwhile, the spatial attention mechanism focuses on learning and emphasizing dense point cloud regions with strong feature expression ability in order to alleviate the influence of uneven LiDAR point cloud density on the results. Finally, the channel attention mechanism is introduced for extracting deeper features after two-branch feature fusion. Experimental verification using HSI and LiDAR data collected by UAV demonstrates that the proposed method outperforms random forest as well as five deep learning methods, yielding more suitable classification results for actual land cover while effectively suppressing the "pepper and salt" phenomenon.