Abstract

A limitation of traditional airborne and spaceborne lidar instruments is the inability to provide data products in real time. This challenge is compounded by typical research-driven desires to build ever more complicated lidar sensors, which overlooks the need to provide simple, but timely, data products to operational forecast models. Machine learning techniques using convolution neural networks (CNNs) have been developed and applied to single wavelength (e.g., 1064 nm) data from the airborne Cloud Physics Lidar (CPL) and have shown encouraging results for feature detection at finer resolutions compared to traditional methods, notably during noisy daytime conditions. Current technologies and properly scoped measurement goals, not intended as be-all/end-all research tools, permit designs for miniaturized lidar sensors that can be placed on drones and, ultimately, in constellations of minisats. Use of advanced machine learning techniques for data processing permits generation of real time data products that can be quickly assimilated into predictive models (for air quality and human health) and for generating real-time data products for decision making (such as hazardous plume detection and monitoring).