Meteorological characteristics of fog events in Korean smart cities and machine learning based visibility estimation

To address various urban issues such as fine dust, traffic congestion, and water shortage caused by rapid urbanization, a national pilot Smart City is planned in two Korean cities, Sejong and Busan. As weather data is crucial for improving the environment and operating future transportation while constructing a smart city, preparing for future weather disasters by analyzing the characteristics of various meteorological phenomena in the planned development area is necessary. This study analyzed the fog generation characteristics for the period of 2016–2020 at the automatic weather system sites of the Korea Meteorological Administration in Sejong and Busan, and the characteristics of the meteorological elements during fog were investigated. Additionally, three machine learning based models, including Random Forest (RF) and Deep Neural Network DNN-1 and DNN-2, were constructed for estimating visibility using meteorological input variables. In Sejong, approximately 50 fog days were observed annually for the analysis period, with the highest frequency of these days being during autumn. The fog hours were distributed from nighttime to before sunrise and 94.3% of the fog occurred when the wind speed was less than 1.5 m/s, showing the characteristics of radiative fog. Most of the fog occurred during summer in Busan, and the maximum number of fog days was observed in July. The average wind speed during fog was approximately 2.5 m/s, which was relatively large, suggesting that advection fog was likely to occur. As a result of estimating visibility from the ML-based models, the RF model had the highest R² (0.66 and 0.53) in both regions, but the visibility estimated from the RF model had an over- and under-estimation for short and long visibility ranges, respectively. DNN-1 and DNN-2 models have lower biases. In the detection of fog-possible weather based on the estimated visibility from the models, the best precisions (0.85 and 0.84) and F1-scores (0.76 and 0.74) were from the RF model, but model recall was better in the DNN-1 model. The model recalls for the detection of the thick and dense fog were also better in the DNN models. ML-based models presented reasonable performance but revealed their weaknesses and strengths depending on performance indicators.