Retrieval of Hyperspectral Surface Reflectance Based on Machine Learning

Many methods based on radiative-transfer models and empirical approaches with prior knowledge have been developed for the retrieval of hyperspectral surface reflectance. In this paper, we propose a novel approach for atmospheric correction of hyperspectral images based on machine learning. A support vector machine (SVM) is used for learning to predict the surface reflectance from the preprocessed at-sensor radiance image. The preprocessed spectra of each pixel are considered as the spectral feature and hypercolumn based on convolutional neural networks (CNNs) is utilized for spatial feature extraction. After training, the surface reflectance of images from totally different spatial positions and atmospheric conditions can be quickly predicted with the at-sensor radiance image and the models trained before, and no additional metadata is required. On an AVIRIS hyperspectral data set, the performances of our method, based on spectral and spatial features, respectively, are compared. Furthermore, our method outperforms QUAC, and the retrieved spectra have good agreement with FLAASH and AVIRIS reflectance products.