关键词: 大气能见度, 空域, 频域, 傅里叶变换, 周向谱, 监督学习

Abstract:

Atmospheric measurement not only impacts marine, land, and air transportation and resident trip, but also is a leading indicator of air quality. The existing visibility estimators based on image processing have problems such as constant computational formulae, poor stability and stringent requirement of application environment. Visibility measurement with supervised learning extracted features related to image edge in spatial domain and features of energy distribution in frequency domain to constitute the high-dimensional feature vector directly from observed scene images, and needed no manual object installing and modeling of the observed scene. It trained Support Vector Regression (SVR) model on the samples that were similar to test image and chosen by k-Nearest Neighbor (kNN), dynamically established the learning model between image features and visibility, and hid various impact factors of visibility in the model. The experimental results of natural scene show that the accuracy of the method can be as high as 96.29%, and moreover, it has good stability and real-time and simplicity of operation so that it is propitious for generalization in large scale.

Key words: atmospheric visibility, spatial domain, frenquency domain, Fourier Transform (FT), circumferential-direction spectrum, supervised learning