Leveraging Spatial Context Data to Improve HOLODEC Segmentation Model Performance

Identifying hydrometeor attributes in atmospheric imaging samples is a critical component in translating campaign aerial data into meteorological insight. However, extracting particle attributes, such as a particle’s spatial position and shape, from holographic imaging samples is a time intensive and complex task requiring great human and computational resources. We present an improved neural network segmentation model designed to locate these particles that is more accurate and computationally efficient relative to its predecessor. The improvements to the pipeline include incorporation of the complete holographic field, sensor depth context data, refined evaluation methods for better shape and size estimations, and streamlined training and execution. This yielded concrete improvements in both accuracy and training time on synthetic data sets. Overall these are important steps taken towards the development of a model that can process field campaign data at the same speed that data is taken in for real-time processing.