ALOS AVNIR-2 Prototype Processor for Atmospheric Correction
(1) DLR - German Aerospace Center, Muenchener Str. 20, D - 82234 Wessling, Germany
ESA granted DLR a contract for the development of a prototype processor for imagery from ALOS AVNIR-2 and PRISM. The processing comprises the systematic (level 0) and radiometric correction (level 1a), orthorectification (level 1b), and atmospheric correction (level 2). This contribution presents design considerations and results for the atmospheric correction (restricted to AVNIR-2) while two other DLR contributions discuss the L0, L1 processors.
AVNIR-2 delivers multispectral imagery in four VNIR bands (blue, green, red, NIR) with a spatial resolution of 10 m in the nadir view. The instrument can also collect data from a +/-44° across-track tilt geometry. The objective is the mapping of land surfaces and coastal zones. Atmospheric correction is a very challenging task for this type of sensor, because the usual techniques require many (hyperspectral) VNIR bands or short-wave infrared bands to derive the aerosol optical thickness (AOT) and subsequently the surface reflectance. Similar to other VNIR sensors (Ikonos, Quickbird) the spectral channels are chosen in such a way that atmospheric water vapour plays a minor role.
The prototype processor employs either a fixed water vapour column or selects a typical seasonal value (e.g., summer, tropical, winter, US standard). Radiative transfer calculations are conducted with the MODTRAN4 code, results are stored as look-up tables (LUT). The final step is a resampling of the fine-spectral resolution LUTs with the AVNIR-2 channel filter functions to obtain the sensor-specific LUTs. These are stored in a database comprising a large range of weather conditions, solar and viewing geometries, and ground elevations.
Since the aerosol type cannot be obtained reliably with a few VNIR channels from land imagery, it is fixed as the rural (continental) type. The AOT is computed with an empirical algorithm masking dense dark vegetation with the red and NIR bands and employing a red/NIR correlation. Additionally, water bodies are searched for and masked. If the scene contains water bodies, the surface reflectance in the NIR band is calculated, and if negative reflectances occur, then the visibility is increased (AOT decreased) until the water reflectance values are positive or the maximum number of iterations is exceeded. If the scene contains no vegetation and no water bodies, then the atmospheric correction is conducted with a constant visibility of 23 km. In case of turbid water the usual NIR surface reflectance threshold = 0.02 or 0.03 will not work to mask water pixels and other criteria will presented.
Clouds and saturated pixels will also be flagged, although it is clear that a unique cloud criterion cannot be achieved with a few VNIR channels. For example, misclassifications may occur for cloud and bright sand. Nevertheless, results obtained with the prototype processor are generally satisfactory.