AI Case Study
University of Essex researchers compare machine learning models for improving weather condition classification of images
University of Essex and Birmingham researchers investigate different methods of classifying images to improve on-demand weather forecasting. Improving the accuracy of image classifiers theoretically allows the transportation industry, among others, to adjust to real-time conditions by automating live conditions updates. The researchers find that a ResNet model provides the best accuracy across the different investigated weather types.
"The specific field of computer vision known as weather classification is still evolving and worthy proposals are still being
made to improve the baseline. An opportunity to advance the
field even further lies in the adoption of new techniques and
tools that have been used to improve other fields. To achieve
this, it is necessary to recognise that weather is a phenomenon
that is a complex phenomena where semantic labelling of images
can be a challenging task. A single image can encompass
multiple weathers for instance partly cloudy. To exploit all
the information that can be extracted from a single outdoor
image, considering uncertainty as an opportunity rather than
a problem is an important step." To this end, researchers from the University of Essex and the University of Birmingham tested different methods for classifying images of weather conditions: "The categories of interest for this contribution are sunny,
cloudy, foggy, rainy and snowy images."
R And D
Core Research And Development
"Weather conditions often disrupt the proper functioning
of transportation systems. Present systems either deploy
an array of sensors or use an in-vehicle camera to predict weather
conditions. These solutions have resulted in incremental cost and
limited scope. To ensure smooth operation of all transportation
services in all-weather conditions, a reliable detection system is
necessary to classify weather in wild. The challenges involved in
solving this problem is that weather conditions are diverse in
nature and there is an absence of discriminate features among
various weather conditions."
"The overall best classification was the ResNet 50 architecture for all of the settings of the superpixel masks. The three variations of the residual networks (ResNet) used in the experimentation were among the four top performing models in the four different settings. These results are a good indication that the optimisation achieved by the inclusion of residual methods learning with shortcut connections has a positive effect in the overall task of weather classification, and it can benefit slightly from the use of superpixel masks as data augmentation. Residual Networks are thus not only successful in arbitrary image recognition, but can also work well as a solution to weather classification problem."
"The pipeline used for the work described in this paper is
based on the work by Kalliatakis et al... There are two
possible paths in the pipeline, either using superpixel masks or
using the raw images directly. In the first mode of operation,
the images are passed through a module that calculates the
superpixels, generates the mask in a colour and applies it over
the image... The images, enhanced with supeprixel mask if used, are divided into four partitions corresponding to positive training, positive testing, negative training, and negative testing. Then, the matdeeprep function, which uses pre-trained Convolutional Neural Network models, is used to extract deep features from the images in all the partitions. Finally, the features are used to train and test a
binary Support Vector Machine classifier."
"For the sunny and cloudy categories, Lu et al.  developed a dataset as a part of their work. This dataset contains images from other datasets and images available online and labelled as part of
their contribution. This dataset was made available on the web
page associated with their paper. The images for the other
categories are those belonging to RFS weather dataset. Since
RFS weather Dataset contains 1100 images for each of its
categories, 1100 images were taken from the categories in Lu
et al. dataset, having a total of 5500 images distributed in five
categories. To create a partition of the images for the Support
Vector Machine classifier training and testing process, 70% of
the images are used for the training phase, while 30% is used
for the testing phase."