Music: a New Hybrid Atmospheric Correction Technique for Solar Reflective Hyperspectral Imagery

Multispectral and hyperspectral imagery usually require atmospheric correction to remove the influence of the environment on the measurements. The images may then be converted in reflectance units. There are generally two different classes of algorithms that attempt to transform at-sensor radiance or raw images into reflectance. One class consists of first principles methods that rely on physical knowledge of how the environment interacts with the surface to produce radiance. These are sometimes referred to as model-based methods or radiative transfer methods. The other class consists of empirical methods which are usually variants of the empirical line methods (ELM) or other in-scene methods. Both classes of algorithms have benefits and drawbacks. The radiative transfer methods such as FLAASH, 6S and ATCOR4 [1,2,3] can achieve high accuracy but require sophisticated computer codes that are computationally time consuming, require accurate characterization of the atmosphere, and generally do not perform well in less than ideal environmental conditions (overcast, dust) or when the sensors characteristics are not accurately known. The Empirical Line Method based algorithms are computationally fast, can be very accurate, and do not rely on well calibrated sensor but since they require a-priori knowledge of in-scene elements, they are not suitable for all applications. The QUAC algorithm [4], that we may call an empirical-statistical approach, eliminates the main drawbacks of the ELM algorithm by not requiring apriori knowledge of in-scene elements. Instead, it relies on an empirical observation that the mean of many scene endmembers (assuming that the scene is diverse enough) is constant and known a priori. In a more recent paper the QUAC authors [5] reported that they had to make some adjustment to the algorithm (excluding some additional types of background) because their assumption about the mean of scene end-members being constant was not always holding. Even with adjustments, algorithms like QUAC that rely very heavily on statistical hypotheses may fail on the inevitable statistical exceptions.