The Visibility Measurement Based on Convolutional Neural Network

Conventional visibility detection instrument has an expensive price and requires for the deployment field strictly. To achieve the goal of low-cost visibility detection, this article studies a low-cost approach which combines monitor camera device and deep learning algorithm. Cameras mounted in the external field are used to capture images, which are the input images of the convolutional neural network (CNN). The scattering visibility detection instrument is used to capture visibility values, which are transferred to the training labels of CNN. Both devices are mounted at the same site outside the meteorological observation station at our school. In advance, 3300 images are filtered from 40000 original sampling data sets to construct a dataset, where 3000 samples are selected as the training set and the rest are as the test set. We don't adopt the common 7:3 dataset partition strategy, because this 10:1 strategy can promise better training performance in the case of a small-sized dataset. This article constructs the network based on the convolutional neural network of Alex model. During training, two methods are employed for preventing over-fitting. Although there is some fluctuation in the accuracy curve during training, the prediction trend is increasing and stabilized to nearly 85% finally. The result curve comparison shows that in this same period the prediction curve approximately coincides with the detected visibility values, and has a similar variation trend. At last, it is concluded that the low-cost visibility detection method based on cameras and CNN is practicable with relatively high performance, which can comply with the application requirements. The following research will focus on improving the accuracy and decrease the size of CNN.