High-emitting vehicle identification by on-road emission remote sensing with scarce positive labels

Abstract

On-road emission remote sensing (OERS) is an ideal means to identify the on-road high-emitting vehicles, which can scan thousands of vehicles within a day without interfering the normal driving. Due to the complex and varying measuring environments and vehicular operating states, it is reasonable to determine the high-emitters not only by the OERS-output pollutant concentration, but also the other information, such as meteorological and vehicular conditions. This paper aims to establish a high-emitter identification model by machine learning technologies to combine the OERS outputs and periodic emission inspection results. The periodic emission inspection, which is conducted in vehicular inspection stations (VIS), is relatively accurate since the measuring environments and vehicular operating states are controllable, and thereby the periodic emission inspection results are considered as the truth values (or labels). However, VIS is extremely inefficient compared with OERS, thus resulting in scarce labels. Moreover, due to some practical issues, such as staff cheating, only the positive labels (high-emitters) are reliable. Therefore, this paper studies the possibility of employing the one-class classification and graph-based label propagation to solve the problem of scarce positive labels. The experimental results show that the high-emitter identification model based on one-class classification can achieve satisfactory performance, which could be further improved by the application of graph-based label propagation.