A substantial proportion of roads, especially in developing countries, remain unsealed. Monitoring of traffic-induced dust is an essential part of unsealed road maintenance. The measurement of the degree of severity of dust emissions is a challenging task due to dust's spatio-temporal characteristics, the requirement for expensive equipment, and rapidly changing environmental conditions. The next evolution of mobility solutions for unsealed roads calls for more robust, time-efficient and smart methods, such as machine learning techniques to solve dust-related road problems. This paper investigates the performance of current semantic segmentation machine learning models on the identification of dust clouds using the previously published URDE dataset. Based on static image performance, the selected models are then used to identify and segment dust clouds in video recordings of road dust emissions. These are used to visually analyse the segmentation quality of the best-performing machine learning models and determine whether the development of novel machine learning models for segmentation of dust is necessary. In this study, we provide code and validation data including segmented videos which will assist in real-world applications.