Estimating Statewide Atmospheric Visibility From Camera Images

Atmospheric visibility is a key meteorological parameter in transportation safety and weather phenomena classification. One promising method to measure atmospheric visibility is utilizing camera images. Wilson Reyes (2023) created a deep neural network model that estimates the atmospheric visibility of a location given a camera image. The model is trained to compare a query image with a reference image, and determine the extent to which the query is less visible than the reference. Once trained, the model can be used to compare a query image to a small set of reference images with known visibility distances. These comparisons are then used to estimate the query image’s visibility distance.

Accurate distance estimates rely on high-quality reference images with calibrated distance labels. We propose a human verification approach in order to maintain the accuracy of the model and to improve end-user trust in the resulting estimates. The process starts with identifying the distance of distinctly visible features from the camera. A large set of images is labeled preliminarily with visibility distances from the nearest Automated Surface Observing System (ASOS) station. Because of the potentially large distance from the ASOS station to the camera, these labels can be very noisy. For each distance range, a domain expert obtains a random sample of images and selects those images whose visibility matches the ASOS-derived distance label. If a feature is faintly visible, then the visibility is considered to be the distance of the feature. If a feature is clearly visible, then the visibility is assumed to be at least one mile greater than the distance to the feature. And, if a feature is not visible, then the image’s visibility is less than the distance to the feature. We used this procedure to create a set of reference images for 115 New York State Mesonet (NYSM) stations (approximately 25-30 images per station).

Given query images from each of the cameras taken at a given time, we then use the trained model to estimate atmospheric visibility across all of the Mesonet stations in the State of New York. These visibility measures are then displayed on a state map. With this approach, we performed three multi-day case studies: wildfire smoke, a nor'easter, and clear conditions. The results show that the model responded well to the stations’ visibility conditions across the three cases.