Detecting the Presence of Precipitation in New York State Mesonet Imagery at Night using Convolutional Neural Networks
A variety of pre-defined convolutional neural network architectures are harnessed to predict the presence of precipitation in New York State Mesonet (NYSM) imagery at night.
A total of 126 stations across the state record panoramic images with a standard temporal frequency of five minutes. Tens of thousands of images, each with diverse backgrounds, are labeled based on the presence or absence of precipitation streaks from infrared detection.
The two main categories of precipitation or no precipitation are expanded upon to predict obstructions such as spider web formation or lens-covering raindrops.
Procedures to enhance trustworthy output are integrated within the image labeling process through inter-coder reliability metrics, and uncertainty quantifications are recorded through probabilistic model predictions.
Unique model architectures are compared and performance metrics are evaluated to deduce model generalizability on future imagery or the geographic expansion of imagery.
The end goal is to create a deep neural network to provide real-time, operational capacities for the prediction of precipitation across New York State to be used alongside the suite of NYSM instrumentation for actionable insight.