Automatic atmospheric correction for shortwave hyperspectral remote sensing data using a time-dependent deep neural network
Automatic atmospheric correction for shortwave hyperspectral remote sensing data using a time-dependent deep neural network
复制标题
DOI：
10.1016/j.isprsjprs.2021.02.007
复制
发表时间：
2021-02-24
影响因子：
12.7
通讯作者：
Salvador, Mark
中科院分区：
文献类型：
--
作者：
Sun, Jian;Xu, Fangcao;Salvador, Mark
大气校正是高光谱成像和光谱仪遥感数据目标检测的重要步骤。最先进的大气校正方法要么需要大量的现场实验，要么需要大气特征的先验知识来提高预测精度，这些方法计算量大且不适合实时应用。为了充分发挥遥感观测快速、可靠获取大面积数据的优势，需要一种自动化、高效的大气校正处理工具。在本文中，我们提出了一种基于时间的神经网络，用于使用不同仰角下的多扫描高光谱数据进行自动大气校正和目标检测。除了总辐射率之外，还纳入了收集日期和时间，以提高网络的时间依赖性，并代表大气和太阳辐射的季节和昼夜特征。结果表明，所提出的网络能够准确提供大气特征并估计精确的反射率光谱，对于植被、海冰和海洋等不同物质，平均准确度为 95.72%。其他实验旨在研究网络对缺失数据的时间依赖性和性能。误差分析证实，我们提出的网络能够估计季节性和昼夜变化环境下的大气特征，并处理丢失数据的影响。预测结果和误差分析都很有前景，并证明我们的网络有能力提供准确的实时大气校正和目标检测。
Atmospheric correction is an essential step in hyperspectral imaging and target detection from spectrometer remote sensing data. State-of-the-art atmospheric correction approaches either require extensive filed experiments or prior knowledge of atmospheric characteristics to improve the predicted accuracy, which are computational expensive and unsuitable for real time application. To take full advantages of remote sensing observation in quickly and reliably acquiring data for a large area, an automatic and efficient processing tool is required for atmospheric correction. In this paper, we propose a time-dependent neural network for automatic atmospheric correction and target detection using multi-scan hyperspectral data under different elevation angles. In addition to the total radiance, the collection day and time are also incorporated to improve the time-dependency of the network and represent the seasonal and diurnal characteristics of atmosphere and solar radiation. Results show that the proposed network has the capacity to accurately provide atmospheric characteristics and estimate precise reflectivity spectra with 95.72% averaged accuracy for different materials, including vegetation, sea ice, and ocean. Additional experiments are designed to investigate the network's temporal dependency and performance on missing data. The error analysis confirms that our proposed network is capable of estimating atmospheric characteristics under both seasonally and diurnally varying environments and handling the influence of missing data. Both the predicted results and error analysis are promising and demonstrate that our network has the ability of providing accurate atmospheric correction and target detection in real time.