5:15 PM - 6:30 PM
[ACG43-P06] Fundamental Study on Weather Judgement of Image Data using Machine Learning Classification
Keywords:machine learning multi-class classification, judgement of weather conditions in image data
The sky and ground conditions in many image data, i.e., pictures and video images, are valuable weather information. Using the multi-class classification technique of machine learning, this study investigated the potential how well the technique judged the weather condition involved in image data. For this purpose, about 30,000 pictures were taken with a 10-minute interval (144 pictures per day) by a field monitoring camera fixed in the Kumamoto University. Corresponding weather condition was then assigned from the published data of Kumamoto weather station. At this timing, the weather conditions were summarized into three categories, i.e., sunny, cloudy, and rainy conditions. The supervised training, validation, and test data were prepared from these pictures and weather condition data. This study employed simple multilayer perceptron with two hidden layers for building a classification model, and investigated the accuracy of the three-category classification of following four cases: (1) all pictures, (2) daytime, (3) nighttime pictures and (4) pictures around sunrise and sunset times. The results showed that the accuracy rates of four cases all reached about 85%. It is interesting that our classification model could judge weather conditions even using nighttime and around sunrise and sunset pictures. However, we confirmed that the significant difference appeared the time until when the accuracy rate reached 80% in the training process. That is, the training of the nighttime pictures needed longer time than the daytime ones. It was suggested that our model might have trained using relative brightness of pictures without convolution.