| dc.description.abstract | Dense fog conditions can result in transportation disruptions when motorists encounter significant reductions in visibility. Caution needs to be heeded when these conditions are experienced an early-warning systems are one of the best defenses for alerting motorists of these conditions. With the prevalence of surveillance cameras along highways to monitor traffic flow, image data is plentiful as these cameras operate 24/7. With access to this data, an automated fog detection system has been developed by KNMI that employs a Convolutional Neural Network (CNN) to classify these images as 'fog' or 'no fog'. While the system has been performing satisfactorily with images during daylight hours, the system is not as robust when it attempts to classify images during nighttime hours.
Consequently, the aim of this study is to develop the architecture of a CNN that has satisfactory performance classifying visibility conditions during nighttime hours. One of the main challenges with this study was the lack of 'fog' image data as dense fog is a fairly rare event. Therefore, two approaches were used in this study. The first was to include some basic data augmentation techniques such as cropping, resizing, rotating, and flipping images to synthetically increase the dataset. The second approach included implementing transfer learning to improve the performance of the classifier. A variety of well-known and
high-performing architectures were employed including the use of the VGG16 model. Since these models were trained on large image data sets, they can be re-used since the features learned can be transferred to a new domain.
The VGG16 model emerged as the best performing model from the four pre-trained models used in this task. Its performance was very similar to the existing model developed by KNMI. Areas of improvement are need in decreasing the number of false positives as many images that were? identified as fog were not labeled as such. Based on CNN architectures used in other weather classification tasks, such as classifying clouds, simpler, less complex network architectures seem to perform the best on this image data set.
While the strategies adopted for use in this study did not produce a classifier with better performance than KNMI's exiting model, the results indicate that this is a challenging problem to solve and underscores the need for continued research in this area. Since one of the main challenges of this problem is the highly skewed image data set to the negative class, artificially increasing the image data set by employing a Generative Adversarial Network may help the network learn more features of nighttime fog images. | |
