Advanced Remote Sensing, Inc.,
Keywords: atmospheric correction, imaging satellite, reflectance based, artificial neural network
Summary:Explosive growth is occurring in the use of satellite imagery, especially for defense uses that include global reconnaissance now under development at the US Department of Defense. When operational, these uses will generate many hundreds of daily images that must undergo AI-based analysis for threat detection and monitoring. While this is the vision, a serious impediment exists in the form of haze caused by light transmission through variable atmospheric content of suspended particles and gasses. Atmospheric correction is a necessary goal, but remains a problem because existing radiance-based methods are imperfect and cannot be made near real-time due to their iterative structure. Through National Science Foundation SBIR funding, Advanced Remote Sensing, Inc. (ARSI) has discovered that atmospherically-induced changes in reflectance, the ratio of radiance to the total incident solar radiation, are highly structured. ARSI translated this structure into a closed form relationship based on observation of atmospheric transmission effects. Rather than radiance, patent pending “Simplified Method for Atmospheric Correction” (SMAC) uses reflectance, a simplification permitting image correction from scene statistics. The result is a method more accurate, across twice the range of atmospheric effects, and in one-tenth the time. While SMAC is potentially a revolutionary solution, it can be greatly enhanced through inclusion of artificial neural network (ANN) technology. The advancement enabling SMAC’s utility is a simple closed-form equation using only two inputs, the slope and offset of variable linear correction across the reflectance continuum from dark to bright for every satellite band. Linear relationships for the correction are determined by assessing the degree of atmospheric effect in the most complex process of SMAC image correction. No part of SMAC is “black box,” and this simplifies identifying only a few inputs to determine the processing workflow to be differentially applied to discrete portions of each image. By identifying few, highly influential variables to assess the correction mode, ANN can simplify the process of differential spatial correction across the image; a task now being approached through complex decision trees. ANN can make SMAC image correction robust and sub-minute for all conditions to be encountered globally. ARSI used Sentinel 2 imagery as the test bed for SMAC development that resulted in highly calibrated relationships to drive image correction. These relationships form families of curves for sensor bands that have systematic, gradual changes of shape according to the wavelengths of each band. ARSI’s next-generation concept is “Single Overpass Calibration” (SOC) that will employ a highly-engineered calibration target design. One to several overpasses of the target by a new satellite provides points for each band to project into the family of curves. SOC provides a mathematical guide for interpolation of the calibration across all levels of reflectance that likely relieves the necessity for inter-satellite radiance harmonization. Additional overpasses of the target can then continue to provide QA/QC. The combination of calibration target and the software to be developed is expected to permit continual surface reflectance correction for hundreds of on-the-fly satellites.