气溶胶光学厚度估测中的LASSO特征选择方法

气溶胶光学厚度估测中通常利用遥感信息构造的多种特征属性作为输入，然而，这些属性中常常存在数据噪音、相互关联性和缺失值，从而降低了估测精度和估测强健性。针对这个问题，基于最小绝对收缩和选择算子（least absolute shrinkage selection operator，LASSO）方法和气溶胶光学厚度反演的先验知识，提出了一种针对遥感卫星观测的高维数据进行特征选择的方法，利用2009年4月2日至2011年4月1日2 a间与全球197个气溶胶地基自动观测网站点时空同步的MODIS（moderate-resolution imaging spectroradiometer）遥感数据，采用常用的人工神经网络作为估测模型进行实验分析，表明该方法能结合反演先验知识对多种异质遥感属性进行分组，通过组间迭代保留关键特征，去除冗余属性，有效进行特征选择，从而显著提高气溶胶光学厚度的估测精度。

A Feature Selection Approach via LASSO for Aerosol Optical Thickness Estimation

For estimating aerosol optical thickness, many constructed attributes from remote sensing data generally are used as inputs for a regression model. However, there might be noisy and correlated attributes in heterogeneous high dimensional inputs. They lead to the loss of estimation accuracy and reduction of robustness. To solve the problem, this paper proposed a feature selection approach by integrating Least Absolute Shrinkage Selection Operator (LASSO) and remote sensing priori knowledge. We validated our approach based on the spatial temporal synchronization data between MODIS and 197 global observation sites in the aerosol robotic network during April 2, 2009 and April 1, 2011. Back propagation neural networks are used as regression models. The experimental results showed that the proposed feature selection approach can group remote sensing attributes according to remote sensing knowledge and effectively select informative features from groups gradually by an iterative LASSO procedure. In this way, the proposed approach significantly improves the estimation accuracy of aerosol optical thickness.