Fusion of optical, radar and waveform LiDAR observations for land cover classification
Fusion of optical, radar and waveform LiDAR observations for land cover classification
复制标题
DOI：
10.1016/j.isprsjprs.2022.03.010
复制
发表时间：
2022-03-17
影响因子：
12.7
通讯作者：
Mountrakis, Giorgos
中科院分区：
文献类型：
--
作者：
Jin, Huiran;Mountrakis, Giorgos
土地复盖是确定人类活动特征和促进可持续土地利用的一个组成部分。在广泛的时空尺度上绘制土地复盖的分布和复盖范围图在很大程度上取决于遥感数据的分类。尽管近年来多源数据融合在土地覆盖分类中发挥了越来越积极的作用，但我们对现有研究的深入回顾表明，光学、合成孔径雷达(SAR)和光探测和测距(LiDAR)观测的集成还没有得到彻底的评估。在这项研究中，我们通过以下方式弥合了这一差距：i)总结相关融合研究并评估其报告的精度改进；ii)进行我们自己的案例研究，其中首次使用空间或适当模拟平台收集的数据来评估光学、雷达和波形LiDAR观测的融合以及分类精度的相关改进。研究了多时相陆地观测卫星5号/专题映射器(TM)和高级陆地观测卫星1号/相控阵L波段合成孔径雷达(ALOS-1/PALSAR)图像，这些图像是在靠近机载波形图(土地、植被和冰传感器)数据的纽约中心地区获取的。分类采用随机森林算法，以传感器和季节性的不同特征集作为输入变量。结果表明，组合光谱、散射和垂直结构信息对不同土地覆盖类型的区分能力最强，总体精度最高，达到83%(比双传感器和单传感器方案分别提高2-19%和9-35%，总体精度分别为-81%和48%-74%)。与PALSAR特征相比，多时相陆地卫星图像与基于LVIS的冠层高度度量相结合取得了更大的改善，这表明LVIS提供了更多有用的专题信息，补充了光谱数据，并有利于分类任务，特别是对植被类。全球生态系统动力学调查(GEDI)是最近发射的一种激光雷达仪器，其性能类似于目前在国际空间站(ISS)上运行的LVIS传感器，我们希望这项研究将作为文献综述，并为进一步应用多日期和多种类型的遥感数据融合改进土地覆盖分类提供指导。
Land cover is an integral component for characterizing anthropogenic activity and promoting sustainable land use. Mapping distribution and coverage of land cover at broad spatiotemporal scales largely relies on classification of remotely sensed data. Although recently multi-source data fusion has been playing an increasingly active role in land cover classification, our intensive review of current studies shows that the integration of optical, synthetic aperture radar (SAR) and light detection and ranging (LiDAR) observations has not been thoroughly evaluated. In this research, we bridged this gap by i) summarizing related fusion studies and assessing their reported accuracy improvements, and ii) conducting our own case study where for the first time fusion of optical, radar and waveform LiDAR observations and the associated improvements in classification accuracy are assessed using data collected by spaceborne or appropriately simulated platforms in the LiDAR case. Multitemporal Landsat-5/Thematic Mapper (TM) and Advanced Land Observing Satellite-1/ Phased Array type L-band SAR (ALOS-1/PALSAR) imagery acquired in the Central New York (CNY) region close to the collection of airborne waveform LVIS (Land, Vegetation, and Ice Sensor) data were examined. Classification was conducted using a random forest algorithm and different feature sets in terms of sensor and seasonality as input variables. Results indicate that the combined spectral, scattering and vertical structural information provided the maximum discriminative capability among different land cover types, giving rise to the highest overall accuracy of 83% (2-19% and 9-35% superior to the two-sensor and single-sensor scenarios with overall accuracies of 64-81% and 48-74%, respectively). Greater improvement was achieved when combining multitemporal Landsat images with LVIS-derived canopy height metrics as opposed to PALSAR features, suggesting that LVIS contributed more useful thematic information complementary to spectral data and beneficial to the classification task, especially for vegetation classes. With the Global Ecosystem Dynamics Investigation (GEDI), a recently launched LiDAR instrument of similar properties to the LVIS sensor now operating onboard the International Space Station (ISS), it is our hope that this research will act as a literature summary and offer guidelines for further applications of multi-date and multi-type remotely sensed data fusion for improved land cover classification.