Creating Convolutional Neural Networks to Monitor New York State Mesonet Daytime Precipitation from Camera Images

The New York State Mesonet (NYSM) standard sites capture images at a five-minute temporal frequency in 126 locations across New York State. The National Weather Service uses these cameras to monitor precipitation, especially in less populous areas. While these can be monitored via the camera viewer on the NYSM website, this requires human attention and many man hours to monitor. A Convolutional Neural Network (CNN) machine learning algorithm is used as an image classification tool to automatically monitor these stations to reduce human load. This project uses Explainable Artificial Intelligence (XAI) methods of labeling codebooks and inter-coder reliability to create datasets of labeled images for NYSM sites, which are co-located with ASOS sites. CNNs from multiple specific NYSM sites will be shown, as well as a generalized CNN for all labeled sites. Initial site-specific results, which train on one camera site and validate on a different camera site, have shown training accuracies above 90% but fail to generalize well through validation data on other camera sites. A generalized model is expected to perform with less accuracy, but future work about methods to improve generalizability will be discussed.