Deep spatiotemporal residual early-late fusion network for city region vehicle emission pollution prediction

Abstract

Regulation on the urban vehicle emission has great impact on our daily lives and can protect public health. However, there are sparse emission remote sensing stations in city, and vehicle emission data is both spatial and temporal non-stationary, which is influenced by various internal and external factors, such as spatial dependencies (nearby and distant), temporal dependencies (closeness, period, trend), environments (road network, meteorology, events, traffic flow and POIs). In this paper, we introduce a semi-supervised learning approach with co-training geographical weighted regression model, which aims to construct the historical emission observations with the insufficient stations records. And then we formulate the region emission prediction as a spatiotemporal sequence forecasting problem and propose a deep spatiotemporal residual early-late fusion network based on unique properties of spatiotemporal data, to predict vehicle emissions in each region of the given city. And the residual convolution network is employed to model the temporal properties of region vehicle emissions. Finally, we present experiments with the remote sensing records of Hefei, where the proposed model outperforms the other baselines. This result demonstrates that combining deep spatiotemporal residual early-late fusion network with the semi-supervised geographical weighted regression can predict vehicle emission in each region of city effectively.

Introduction

With the increasing number of urban motor vehicles, the ecological and environmental problems caused by automobile exhaust emissions have become increasingly prominent, and have raised a hot topic of social concern. Vehicles generate greenhouse gases, such as CO₂, HC, NO_x and PM_2.5. Getting real-time vehicle emission information is important to support urban traffic pollution control and protect environment. If we were to know the city region vehicle emission at any time, which can enable region pollution alerts and help relevant governments improve city’s transportation infrastructure design.

For example, there is a pollution source in places where geographical and meteorological conditions are not in favour of ventilation. To alleviate high atmospheric pollution, it is necessary to take strict actions as closing of schools and industries and restrict vehicles circulation. If it were possible to predict the place with high pollution probability one or two days in advance, more efficient actions could be taken to alleviate the potential region pollution [1].

Existing methods for region vehicle emission prediction can roughly be categorized into two classes, namely and classical dispersion models and satellite remote sensing. For the classical dispersion models, such as gaussian plume models, operational street canyon models and computational fluid dynamics. And the mobile source emission factor (MOBILE) and computer programme to calculate emissions from road transport (COPERT) models developed in USA and Europe respectively are the most frequently used emission factor models [2]. David D.Parrish evaluated the estimation of the mass of the emitted species, the temporal evolution of the annual average emissions over decade scales, and the speciation of the VOC emissions [3]. Dane Westerdahl et al. characterized Beijing on-road emission factors, the impact of on-road transportation on air quality, to provide data and control measures in advance of the 2008 Olympics [4]. David R. Lyon et al. construct a spatially resolved methane emissions inventory for the 25-country Barnett Shale region with estimation of emissions from O&G and other sources using bottom-up approaches. These models are usually a complex model of meteorology, road network geometry, geographical locations, traffic volumes and emission factors, based on a number of empirical assumptions and parameters which might not be applicable to all city regions [5]. And these parameters are usually hard to obtain, and the results generated by these kinds of models may be inaccurate. Satellite remote sensing of surface air pollution has been studied intensively in past decades [6], which can be regarded as top-down methods. Donkelaar et al. [7] compared PM_2.5 inferred from the moderate resolution imaging spectroradimeter. Lamsal et al. [8] estimated surface NO₂ concentrations by applying local scaling factors from a global three-dimensional model. Zou et al. [9] employ a geographically weighted regression (GWR) model to predict the urban PM_2.5 using the high-resolution satellite aerosol optical depth. However, such approach is extremely influenced by clouds and would be sensitive to other environmental factors, such as humidity, temperature, pressure and geographical locations [7].

Although much progress has been made by these methods, it is still a challenging problem for the estimation of region vehicle emission in the real-world [10].

Firstly, there are insufficient remote sensing vehicle measurement stations in a city for the expensive cost of building and maintaining such equipments. And urban vehicle emission varies by locations non-linearly and depends on many complex external factors, such as road networks, meteorology, traffic, green land ratio and living functions. Moreover, the region emission distribution is with spatiotemporal dependencies. And we summarize the challenges in the region vehicle emission prediction as following:

1) Data sparsity and spatial heterogeneity [11]. As can be seen from Fig. 1, the blue grids denote the location of the monitor station in Hefei. The vehicle emission monitor stations in city is very sparse. And it is hard to make the emission prediction on the city region based on the limited remote sensing monitor stations. Moreover, the spatial heterogeneity of the Emission-ExternalFactor relationship should be taken into account, or in other words, the strength of the Emission-ExternalFactor correlation should not be constant across space and it should change with spatial context. Most existing researches [12] for spatial interpolation are based on geographical statistics methods, which employ the first law of geography: near things are more related than distant things [13]. In the region emission interpolation problem, these methods neglect the road network topology and the traffic volume influence, which means that the two regions are far away in geographic space but have similar road network topology and traffic volume and they may have similar emission distribution. Using such assumptions may cause unpredictable errors. This motivates us to propose a extended co-training geographical weighted regression model, which combines the road network topology and the traffic effects.

2) Spatiotemporal dependencies. In the spatial scale, the vehicle emission of Region R2 (in Fig. 2) is affected by the emission diffusion of nearby regions (R1 and R3). In the temporal scale, the region vehicle emission is influenced by recent, near and distant time intervals. For instance, a traffic congestion occurring at 10 am will affect that of 11 am. Moreover, traffic conditions in the morning may be similar on consecutive workdays, repeating every 24 h. The traditional air pollution method [14] considers it as a single-point time series prediction problem, and then performs spatial interpolation to achieve overall region prediction. But it does not take spatiotemporal dependence into account. That is, each grid emission is affected by its time and spatial neighbors. Without considering the spatiotemporal dependence, we can not achieve a precise region emission prediction. Therefore, we consider the spatiotemporal dependence in each step of prediction with the spatiotemporal residual network.

All these challenges inspire us to rethink the region vehicle emission prediction problem based on deep learning model with the rich amount of spatiotemporal data [15]. And to address the aforementioned issues, in this paper we predict the region vehicle emission through a data-driven method, using a variety of datasets, including remote sensing records, meteorological data, traffic data, road networks data and POIs. Specifically, we present a semi-supervised learning approach with co-training geographical weighted regression (COGWR) to address the data sparsity and spatial heterogeneity, and formulate the region emission prediction as a spatiotemporal sequence forecasting problem that can be solved by constructing a deep learning framework. In order to model the spatiotemporal relationships, we propose a deep spatiotemporal residual early-late fusion network (ResNetELF) to collectively predict vehicle emission in every region. The effectiveness of the proposed method is demonstrated by the comparison with several baseline methods on the real-world dataset. The main contributions of this paper are as followings:

1) To deal with the data sparsity and spatial heterogeneity of vehicle emission records, we propose a semi-supervised learning approach with co-training geographical weighted regression (COGWR), which leverages available emission data combined with meteorology, traffic, POIs and road network to fill the missing entries at unmeasured locations.

2) To capture the spatiotemporal dependencies of region emission, we proposed an end-to-end deep neural network architecture (ResNetELF), which employs convolution layers to model spatial dependencies while extracting the temporal properties of vehicle emission as closeness, period, and trend dependencies. Moreover, to analysis the impact of external factors, an AutoEncoder module and an early-late fusion strategy are proposed to integrate the extracted spatiotemporal feature with multiple external factors (e.g., traffic, weather conditions, road network).

3) The proposed approach is evaluated on the vehicle emission remote sensing data of Hefei. The results demonstrate the superiority of our proposed approach against the state-of-the-art methods.

The rest parts of this paper are arranged as follows. The related work is summarized in Section 2. The system overview is presented in Section 3. The proposed COGWR algorithm and deep spatiotemporal residual early-late fusion network are clarified in Section 4. Section 5 gives the details of experiment settings and the related experiment results. Finally, we conclude the paper in Section 6.