激光雷达数据协助下的高光谱图像三维残差网络分类

3D Residual Network for Hyperspectral Image Classification Aided by LiDAR Data

摘要: 高光谱成像(Hyperspectral Imaging, HSI)数据可以在狭窄和连续的波段中收集数据，即使在差异很小的情况下，也可以检测到不同类别的数据。然而，光谱变异性和阴影效应会限制任何HSI的分类精度。与HSI相比，激光雷达数据(LiDAR)是HSI的一个很好的补充，因为它提供了高度信息和来自不同物体的多次回波数据。在激光雷达数据的辅助下，通过HSI分类的预处理和深度残差网络分类，可以有效解决阴影效应光谱变异性的问题。除此之外，来源于点云数据的数字高程模型(Digital Elevation Models, DEM)也能够提高HSI分类性能。通过在MUUFL港湾高光谱和激光雷达机载数据集上进行实验的结果表明，采用两种DEM栅格层结合HSI的分类准确率为98.16%，而仅采用HSI数据集独立学习方法的分类准确率为96.3%。并且随着精度的提高，标准偏差从0.304降低到0.150。

Abstract: Hyperspectral imaging (HSI) allows data to be collected in narrow and continuous wavebands, and even when the differences are small, different categories can be detected. However, spectral variability and shadow effects can limit the classification of any hyperspectral image. Compared with HSI, the LiDAR data is an excellent complement to HIS, because it provides both elevation information and multiple returns of echoes from different objects. A procedure including preprocessing and deep residual network classification is investigated for classification of HSI aided by the LiDAR data to solve the problem of identifying shaded objects and spectral variability. In addition, Digital Elevation Models (DEM) derived from point cloud data can also improve the performance of HSI classification. Experiments were performed on the MUUFL bay harbor hyperspectral and LiDAR airborne data sets. The results show that 98.16% classification accuracy was achieved when using two DEM raster layers combined with hyperspectral images, compared to 96.3% accuracy using independent learning methods derived from only the HSI data set. As the accuracy increased, the standard deviation decreased from 0.304 to 0.150. The former indicates that the effect of spectral variability is mitigated.