Improved target detection for hyperspectral images using hybrid in-scene calibration

Abstract

This paper presents a practical approach to target detection for hyperspectral images. In target detection, it is normally assumed that the ground truth target signatures collected in a laboratory are available and one then uses them to search for targets in a given image. However, directly applying the laboratory signatures to the real data is not appropriate due to environmental differences between the ground truth data and real data. Conventional atmospheric compensation schemes such as the use of MODTRAN can help to improve the target detection performance. However, the computational load is huge and thus real-time applications may prohibit this compensation approach. We present results of an alternative compensation technique known as in-scene compensation, which is appealing as no complicated techniques such as MODTRAN are needed. Two in-scene methods for visible near-infrared/short-wave infrared range have been developed in the literature: empirical line method (ELM) and vegetation normalization (VN). Both approaches have advantages and disadvantages. We propose a hybrid in-scene compensation method that can be considered as a combination of ELM and VN and we call our method ELM augmented VN (EAVN). One key advantage of EAVN is that it combines the advantages of ELM and VN and eliminates their disadvantages. Compared to ELM, there is no need for two or more known target pixels in the test scene. Compared to VN, there is no need for dark pixels. Extensive experimental results using ground-based sensor data showed that the EAVN algorithm provides excellent compensation to environmental changes. After compensation, the receiver operating characteristics performance of target detection has been significantly improved by orders of magnitude in a number of cases, as compared to two standard compensation methods: quick atmospheric correction and internal average relative reflectance correction.