Home Index About Next Previous Comments

A Study of Weather Detection in Public Webcams Using Neural Networks Abraham Wong Abstract    This paper describes a framework for detecting weather conditions at outdoor webcam locations. The concept is to train a neural network on webcam, radar, and satellite data. While this goal seems plausible, experiments did not support the approach. Introduction    The Pennsylvania DOT runs a network of traffic cameras in the Pittsburgh area. Detecting weather in the cameras' images could provide near real-time, fine-grained, "usable" weather information, versus standard weather forecasts. Neural networks were selected because of their success in other similar applications. Methods    Download images from National Weather Service and NASA weather sources, and associate points on the images with traffic cameras. Label traffic camera images with the weather data. Then, train a neural network with scaled-down image data, a time component, and the weather labels. These experiments used training data captured over a 6-day period in April. Results    All experiments failed to produce a working detector. Invariably the network would converge to producing the same output regardless of the input image. The networks were almost all 3-layer and feed-forward, and trained using backpropagation. Variables changed included the network parameters (momentum, etc.), input and output formats, image size, synapse type (full, Sanger, or Kohonen). While the NWS radar data was quite accurate for identifying precipitation, the NASA satellite data proved too coarse-grained to accurately determine cloud cover. The satellite data was omitted for most experiments. Possible reasons for the failures include: poor image quality or noise, slight camera movement causing misalignment, and an inadequate network structure or learning parameters. Privacy    The privacy implications of detecting weather in traffic cameras is minimal, but the traffic cameras themselves may uncomfortable to some. While the resolution of PennDOT's public cameras is inadequate to identify faces, it is often sufficient to identify car make, model, and color. Identifying cars could, for example, lead to automated speeding tickets or tracking driving habits by insurance companies. References    Cameras: https://www.nb.net/~finals/allcams.htm Artificial neural network package: Joone Project Paper (PDF)