Modeling the Impact of Weather and Context Data on Transport Mode Choices: A Case Study of GPS Trajectories from Beijing

Abstract

Over the years, researchers have been studying the effects of weather and context data on transport mode choices. Existing research studies are predominantly designed around travel surveys, but the accuracy of their findings relies on how travelers give accurate and honest answers. The proliferation of smartphones, however, now offers the possibility of utilizing GPS positioning data as an alternative information source, opening the potential to accurately model and better understand factors which influence transport mode choices, compared to travel surveys. The objective of this work is to develop a model to predict the transport mode choices based on GPS trajectories, weather and context data. We use 2671 GPS trajectories from the Geolife GPS trajectories dataset, weather data, such as temperature and air quality, and context data, such as rush hour, day/night time and onetime events, such as the Olympics. In the statistical analysis, we apply both descriptive and statistical models, such as the multinomial logit and probit models. We find that temperature has the most prominent effect among weather conditions. For instance, for temperatures greater than 25 °C, the walking share increases by 27%, and the bike share reduces by 21%, which is line with the results from several survey-based studies. In addition, the evidence of government policy on transport regulation is revealed when the air quality becomes hazardous, as people are encouraged to use environmentally friendly transport mode choices, such as the bike instead of the bus or car, which are known CO₂ emitters. Our conclusion is that GPS trajectories can be used as a means to model passenger behavior, e.g. the choice of transport mode, in a quantitative way, which will support transport mode operators and policy makers in their efforts to design and plan the transport mode infrastructure to best suit the passengers’ needs.

1. Introduction

At present, the transport sector accounts for 16% of all world greenhouse gas emissions, and 25% of global carbon dioxide (CO₂) emissions, consequently contributing toward the global climate change [1,2]. Climate change mitigation requires finding ways of achieving sustainable mobility options without compromising economic growth and social inclusion, which necessitates effective government planning and regulation of the transport sector. Therefore, it is crucial to investigate the scientific evidence which shows how changes in climate conditions, transport planning, and transport regulation are intrinsically interrelated.

With climate change, weather is now an important topic in transportation research. Over the years, researchers have been studying the effect of weather and context on the transport mode choice. We define context data as information that can provide perspective into a person or an event, such as culture and habits. These studies relate context information, such as individual characteristics [3], temporal factors [4], transportation supply [5], travel demands [6], and weather conditions [7], to existing or self-gathered travel behavior data.

Most of the works focus on weather effects on transport mode decisions across different cities [7,8,9]. The existing literature on weather and daily mobility focuses particularly on precipitation [10,11,12,13,14], temperature [15,16,17], humidity [18,19], wind speed [19], air quality [17,19], and seasons [20]. A comprehensive description in the study of the influence of weather and context information on transport mode choices is provided in

Table 1. Recent state of the art on the influence of weather and context information on transport mode shares.

Description	Data Source	Travel Modes	Main Variables
Metro ridership fluctuation due to weather [10]	Fare collection system from Nanjing	Metro	Temperature, precipitation, seasonality
Travel behavior changes due to weather [11]	Flemish travel survey		Temperature, rainfall, wind, shopping, leisure trips
Analysis of weather conditions on transit ridership [12]	Automatic vehicle location system from Gipuzkoa	Bus	Precipitation, temperature, wind speed, air humility
Weather and daily mobility in international perspective [13]	Survey data from Utrecht, Oslo, Stavanger, Stockholm	Bike, walk, bus, train, car	Temperature, wind, rain, snow, fog
Impact of weather on urban transit ridership [14]	Fare collection system from NewYork	Subway	Night time, darkness, temperature, rain, snow
Climate change impact on mode choices and traveled distances [15]	Survey data from Randstad	Walk, bike, car, train, bus, tram	Seasons, travelled distance
Machine learning classifiers for modeling travel mode choices [16]	Dutch travel survey	Walk, bike, car, train, bus	Precipitation, temperature, wind, built environment, household income, individual characteristics, distance
Impact of weather conditions on middle school students’ commute mode choices [17]	Beijing school commute Survey	Walk, bike, car, bus, train	Wind, temperature, air quality, humidity
Impact of weather on bus ridership [18]	Smart card data collected in Fengxian	Bus	Humidity, wind speed, rainfall and temperature
Influence of weather on the intercity travel [19]	Survey data from Xi’an	Airplane, high-speed rail, train, bus	Temperature, relative humidity, rainfall, wind, air quality index
Influence of weather and seasonality effects on commute mode choice [20]	Survey data from Chicago	Walk, bike, car	Seasonality
Meteorological variation in daily travel behavior [21]	Dutch national travel household survey	Walk, bike, train, bus, tram	Fog, temperature, precipitation, cloud cover, snow, thunderstorm
Cycling or walking? [22]	Household, personal and travel diary from the Netherlands	Walk, bike	Season, weather, trip characteristics, built environment, work conditions
Weather, transport mode choices and emotional travel experiences [23]	Travel diary from Rotterdam	Walk, cycle, car, bus, train	Temperature, Wind, Sky cleanness, Precipitation
Our study: Modeling the impact of weather and context data on transport mode choices	GPS trajectories from Beijing	Walk, bike, car, bus, train	Temperature, precipitation, relative humidity, wind speed, air quality, rush hours, holidays, day/night, Olympics, trip distance

. The findings of these studies give insights into the separate roles of not only weather parameters, but also the geographical context, cultures and habits in influencing transport mode choices. For instance, the authors in [16,21] indicate that in the Netherlands, higher but not too high temperature favors walking and biking over motorized transport, thus, temperature is more important than precipitation. In contrast, researchers in [10,11] concluded that precipitation creates more significant ridership fluctuations than temperature in Nanjing, China and Flanders, Belgium.

In search for a better understanding of how the geographical context and habits affect mobility decisions, Böcker et al. [13] investigated weather and daily mobility across Dutch, Norwegian and Swedish cities. The authors found that biking is favored by dry and warm conditions, and the use of the car is favored during wet and windy weather. The authors also highlighted the presence of differences in the effects of weather on mobility across the different cities and countries.

With respect to seasonality, Hyland et al. [20] reported that it is highly likely for the car to be the chosen mode of transport during bad weather for commuters in Chicago. In fact, the authors showed that millennials are inclined to choose the car in winter than in the summer, while for non-millennials, seasonality has little impact on the choice of mode of transport. In [22], Ton et al. further highlighted that apart from seasonality and weather characteristics, other categories of context information, such as work conditions, trip and household characteristics, and the built environment, do also influence the transport mode choices. In [23], Böcker et al. investigated how emotional travel experiences influence transport mode choices.

While the aforementioned studies in Table 1 improve our understanding of weather and context data effects on daily mobility choices, one key issue standouts, many of the studies are based on travel surveys or travel diaries for characterizing individual travel behavior. The disadvantage of travel surveys is that the accuracy of their findings relies on how respondents give accurate and honest answers. One alternative source of travel data is to use data recorded by electronic devices and management systems, such as the fare collection system [10], automatic vehicle location [12] and the smart card data systems [18]. These alternatives can provide accurate information of the location, time, date, and transport mode of the traveler than travel surveys. With the continuous proliferation of smartphone devices [24], an opportunity now exists to use satellite positioning data as an alternative source of data in transportation research. To the best of our knowledge, there is no study that analyses the relationship among weather, context information, and different transport modes using satellite positioning data (see data source column in Table 1).

The goal of this paper is to show the possibility of using GPS trajectories in investigating the impact of weather and context data on transport mode choices in Beijing city. As illustrated in Figure 1, not only do we study the impact of weather conditions, such as precipitation, temperature, air quality, wind speed, and relative humidity on transport mode choices, but also the effect of context information, such as rush hour, trip distance, day/night time, holidays, and an event, such as the Olympics, on individual travel behavior is presented. We match data from three different databases as shown in our proposed approach in Figure 1, and apply both descriptive and statistical models such as the multinomial logit (MNL) and probit (MNP) models during the statistical analysis. The specific contribution of this paper is two fold:

• The potential of using GPS trajectories to analyze and model the relationship between transport mode choices, weather and context information is investigated.
• The relationship among weather and context information, transport planning, and transport regulation is analyzed.

Figure 1. Our proposed approach of the analysis of the impact of weather and context data on transport mode choices based on GPS trajectories.

Figure 1. Our proposed approach of the analysis of the impact of weather and context data on transport mode choices based on GPS trajectories.

The structure of the remainder of this paper is as follows. Section 2 presents the databases applied in this work. The statistical analysis and models are described in Section 3 and Section 4, respectively. The results are discussed in Section 5, while the conclusions and future work are presented in Section 6.

2. Databases

The case study is Beijing (see Figure 2), which is the capital of the People’s Republic of China. With a population of over 21 million residents, it is regarded as the most populated capital city in the world. Due to its high population density, rapid urbanization and motorization, Beijing faces severe congestion and air quality problems.

Beijing is an interesting case study due to the sheer number of solutions that have been adopted to overcome its transport related problems, such as the development of bus rapid transit corridors, new extensive bus lanes, and policies, such as congestion charging, among others.

In this section, we describe the three databases with the GPS trajectories, weather and context information in Beijing as proposed by our approach in Figure 1. The benefit of our proposed approach is that now there is a possibility of utilizing the GPS trajectories as an alternative information source, opening the potential to accurately model and better understand factors which influence the transport mode choice behavior, compared to travel surveys.

2.1. GPS Trajectories

The dataset adopted in this study has GPS trajectories of 182 users collected during the Geolife project conducted between April 2007 and August 2012 [25]. The GPS trajectories were recorded by GPS loggers and phones and are described by a sequence of time-stamped points, with each containing longitude, latitude, altitude, and transportation mode label.

Although the dataset contains trajectories distributed in over 30 cities worldwide, in this work, we use a total of 2671 labeled trajectories from Beijing. The total traveled trip distance and duration is 21,350 km and 1296.6 h, respectively. Each user recorded on average 74 trips, i.e., 74 trips is in relation to the average number of dedicated trips recorded by the travelers. Note that not every trip was recorded by the travelers during the five years, but only dedicated trips. The average distance and duration of each trip was 5.75 km and 0.5 h, respectively. In Figure 2, we show the geographical location and demarcation and the starting point of the trips in Beijing.

The transport modes considered in this work include walk, bike, car, bus, and train. The percentage of the transport mode in the dataset is shown in Figure 3. During the period in which the data were collected, walking accounts for 46% of the transport mode labels in Beijing. This result is expected, given that travel guides recommend traveling around Beijing on foot as the best and most efficient commuting mode [26].

It is worth noting that the Geolife trajectory dataset is natural since it was recorded while the users performed their life routines [27]. For instance, trajectories were recorded as users made trips from home to work and vice versa, to entertainment and sports venues, shopping, and sightseeing. During these trips, the users chose a unimodal transport mode, and information about their individual characteristics, household characteristics, trip characteristics, and work conditions was anonymized.

The Geolife trajectory dataset has been used in different research fields such as in privacy-preserving location data [28], measuring trajectory stops and moves [29], user identification [30], trajectory completion [31], and transport mode detection [32].

2.2. Weather Information

The weather conditions considered include, temperature, precipitation, wind speed, air quality, and relative humidity, collected from a meteorological station located within the city of Beijing. This database can be accessed from the National Aeronautics and Space Administration (NASA) website [33].

Thanks to its continental climate, Beijing has hot, sultry, and rainy summers, cold and sunny winters, and a precipitation of about 545 mm annually [34]. However, a summary of the conditions in

Table 2. Descriptive characteristics of continuous observable variables.

	Mean	Std.dev	Min	Max
Temperature (°C)	19.153	8.477	−9.22	38.45
Precipitation (mm/h)	0.083	0.307	0	5
Relative Humidity (%)	53.837	22.076	6.25	100
Wind speed (m/s)	3.064	1.742	0.05	12.37
Air quality (μg/m3)	153.016	64.545	13	547
Distance (km)	7.993	12.341	0	117.216

shows that during the period of study: (i) average temperature was about 19 °C; (ii) the rains were not abundant with a mean of 0.083 mm/h; (iii) the air was less humid with a mean relative humidity of 53%; (iv) the wind speed was very light according to Beaufort scale, given that the mean wind speed was 3 m/s; and (v) the air quality was unhealthy, given that the mean air quality was 153 $\mathsf{\mu }{\mathrm{g}/\mathrm{m}}^3$.

2.3. Context Information

This study also includes the impact of context information, such as the Beijing 2008 summer Olympics event held from 8th to 24th August on transport mode choices. This event was awarded to Beijing in the year 2001, thereafter its leaders embarked on massive projects to transform the city’s transport system. For instance, the rail network was expanded from 50 km to 200 km, and there was an introduction of 286 km of dedicated on-road lanes [35]. In fact, a new set of restrictions were passed in the months leading to the games. For instance, in some parts of the city, people were not allowed to use their private vehicles, trucks from outside Beijing were to avoid the city, and flexible retail and shopping hours were also introduced to spread traffic loads.

Other context information attributes we look at include rush hour or off-rush hour, holiday or non-holidays, day time or night time, and trip distance. In Beijing, morning and evening rush hour traffic demand is from about 7 a.m. to 9 a.m. and 4 p.m. to 8 p.m., respectively, according to the studies made by [35], which was applied during our analysis. As for the holidays, there are 10 public holidays spread throughout the year, which include New Year’s, Chinese New Year, Lunar New Year, Qingming Festival, Labor Day, Dragon Boat Festival, Mid-Autumn Festival, Golden Week, and Christmas [36]. In total, our analysis considers 28 days of the year as public holidays. Day time in Beijing is considered as the period between sunrise, i.e., 6 a.m., and sunset, i.e., 6 p.m., compared to night time, which is between sunset and sunrise.

2.4. Matching Weather and Context Information with GPS Trajectories

As shown in our proposed approach in Figure 1, the GPS trajectories from the Geolife dataset contain the location, date, time, and the mode of transport chosen by the traveler. The starting point of the recorded trajectory provides the departure information of the traveler, such as their departure location, date, and time.

In matching the weather data to the GPS trajectories, the starting point of each trajectory is linked to the hourly historical weather data from the nearest meteorological station. The result generates weather-related variables corresponding to the departure time of each trip. It is worth noting that hourly weather data are preferred over daily weather data because the former generates higher temporal accuracy, given the constantly changing weather.

Similarly, the timing corresponding to each of context information attributes is linked to the departure time of each trip. From the combined datasets, we analyze the impact of weather and context data on transport mode choices by considering weather and context information as input and modal choices as output conditions.

3. Statistical Analysis

In this section, the weather conditions, context information and GPS trajectories are analyzed statistically. We apply two main statistical methods which are commonly used in data analysis: (i) descriptive statistics, where we summarize the findings using bar graphs, and (ii) statistical interference, where we apply the MNL and MNP models.

In order to track the weather changes, we classify the weather conditions into different levels according to their rating generated from published weather knowledge. For instance, the temperature scales with seven levels observed in Figure 4a are adopted from [37,38], precipitation with five levels in Figure 4b from [39,40], wind speed with six levels of classification in Figure 4c are adopted according to the Beaufort number [41], air quality yardstick with six levels that run from 0 to 500 in Figure 4d adopted from monitoring the fine particulate matter concentrations (PM_2.5) [42], and relative humidity in Figure 4e falls into four levels adopted from [43]. These classifications as well as the timing corresponding to the context information are used to compute the share for each transport mode in Figure 4 and Figure 5 due to the effects of weather and context data, respectively.

3.1. Influence of the Weather Condition on the Transport Mode Choices

In Figure 4a, it is observed that during snowfall, cars have a higher share compared to other transport modes. However, walking becomes the most preferred transport as the temperature increases. At temperatures greater than 25 °C, the bike and walking shares reduce and the bus share increases. In Figure 4b, we see that the share of the transport modes is similar across the precipitation levels, except when precipitation is greater than 2 mm/h, i.e., little rain hardly impacts passengers’ travel choices, which is in line with the observations of Junlong Li et al. [10].

In Figure 4c, it observed that as the intensity of the wind increases from calm, greater than 0.27 m/s to strong wind, greater than 7.8 m/s, the bike and walk share reduce, while the car share increases. In Figure 4d, it is noticeable that as the air quality turns from good (0–50) $\mathsf{\mu }{\mathrm{g}/\mathrm{m}}^3$ to unhealthy (201–300) $\mathsf{\mu }{\mathrm{g}/\mathrm{m}}^3$, the transport mode share remains constant. The potential explanation is that Beijing averages an air quality of 150 $\mathsf{\mu }{\mathrm{g}/\mathrm{m}}^3$ (see Table 2), so it is highly likely that most of the time, the air quality is unhealthy as people carry on with their daily lives. However, when the air quality becomes hazardous (300+) μg/m³, bike and walk have a larger share, and there is a reduction in the car and bus shares in order to reduce the (CO₂) emissions. In Figure 4e, the share of transport mode is the same regardless of the relative humidity, which is in line with the findings of Aultman-Hall et al. [38], who found that humidity has a limited effect on transport choices.

3.2. Influence of Context Information on the Transport Mode Choices

During the Olympics, an increase in the bus share and a reduction in the car share occurred as shown in Figure 5a. It is intuitive that during international events of this magnitude, some roads are closed off or restricted to the public. Our findings show that the alternative travel means is by public transport, the bus in particular. In Figure 5b, an increase in the bus share and a reduction in the train and car shares are noticeable at rush hours in comparison with non-rush hours. This finding is in line with the study made by the Asian development bank in [35], in which it was shown that during rush hours, passengers prefer taking the bus over the train.

In Figure 5c, we see that there is an increase in the car share and a reduction in the bus share during holidays perhaps because people are choosing to travel by car on holidays. In Figure 5e, it can be observed that trip distance has a tremendous influence on the transport mode choice. For instance, for shorter distances (0–6 km), walk and bike have a larger share but in between (6–12 km), the bus share is larger. However, between 12–20 km, the bus share reduces as the car and train shares increase. Beyond 20 km, the train share reduces and the car share increases.

4. Statistical Modeling

Given that the transport mode choices are a typical example of discrete outcomes, in order to link the probabilities of choosing a given mode of transport to the weather and context information, in this study, we adopt MNL and MNP models. These are among the most popular models in analyzing and predicting travel decisions. It is worth noting that the graphs shown during the statistical analysis can only be used to identify relationships and patterns, but cannot be used to draw conclusions about the travel behavior of Beijing’s population. This fact motivates the need for the statistical models.

The models analyze simultaneous effects of weather and context information on transport decisions, including the onetime event Olympics in one integrated model. As suggested by Böcker et al. [13], this approach is important in understanding mobility decisions.

4.1. Methods

According to Ben-Akiva et al. [44], the framework of these models is based on four general assumptions: (i) decision maker who is the individual or group of people making the choice, (ii) alternatives, which are the available choices to choose from, (iii) attributes, which are parameters that characterize each alternative, and (iv) decision rule, which is the process used by the decision maker evaluates the alternatives.

The decision rule in models used in travel behavior analysis is based on the utility model, $U_{in}$, and takes the form

$$U_{in}=V_{in}+{ϵ}_{in}$$

(1)

The alternative with the highest utility is chosen; therefore, the probability that alternative i is chosen by an individual n from a choice set $C_n=\{1,2,...,i,...,j_n\}$ takes on the general form

$$\begin{array}{cc}\hfill P_{in}& =P[V_{in}+{ϵ}_{in}⩾V_{jn}+{ϵ}_{jn},\forall j\in C_n]\hfill \\ \hfill & =P[(V_{in}+{ϵ}_{in})=ma{x}_{j\in C_n}(V_{jn}+{ϵ}_{jn})]\hfill \\ \hfill & =P[{ϵ}_{jn}-{ϵ}_{in}\le V_{in}-V_{jn},\forall j\in C_n]\hfill \end{array}$$

(2)

$V_{in}$ is the deterministic or systematic part of the utility function. It is defined in (3) by a vector of observable variables $z_n$ and their corresponding coefficients ${\gamma }_i$:

$$V_{in}={\gamma }_i^T{z}_n$$

(3)

To identify the model, one set of coefficients of the systematic term needs to be normalized to zero, e.g., $({\gamma }_1=0)$, which makes the corresponding transport mode choice $(i=1)$ become the base mode. The coefficients of other alternatives are interpreted in reference to the base outcome.

${ϵ}_{in}$ is the random term which expresses the errors of the utility function. Its distribution is often known and makes the problem more reasonable to empirically characterize. If the error terms assumes an independent and identically distributed ”Extreme Value”, i.e., Gumbel, the model is MNL; hence $P_{in}$ takes on the form

$$\begin{array}{cc}\hfill P_{in}& =\frac{e^{V_{in}}}{{\sum }_{j\in C_n}{e}^{V_{jn}}}\hfill \end{array}$$

(4)

However, when the error terms assume a multivariate normal distribution, the model is MNP and $P_{in}$ takes on the form

$$\begin{array}{c}\hfill {\int }_{{\xi }_{1n}}^{\infty }...{\int }_{{\xi }_{(i-1)n}}^{\infty }{\int }_{{\xi }_{(i+1)n}}^{\infty }...{\int }_{{\xi }_{jn}}^{\infty }\varphi ({\widetilde{ϵ}}_{1n},...,{\widetilde{ϵ}}_{(i-1)n},{\widetilde{ϵ}}_{(i+1)n},...,{\widetilde{ϵ}}_{jn})d{\widetilde{ϵ}}_{1n}...d{\widetilde{ϵ}}_{(i-1)n}d{\widetilde{ϵ}}_{(i+1)n}...d{\widetilde{ϵ}}_{jn}\end{array}$$

(5)

where $\varphi \left(\right)$ is the density function, which follows a multivariate normal distribution with means 0 and a covariance matrix with a size of $J_{n}X{J}_n$, and ${\xi }_{1n}=V_{in}-V_{jn}$ and ${\widetilde{ϵ}}_{in}={ϵ}_{jn}-{ϵ}_{in}$, as also seen in (2), are the difference between the systematic and error terms, respectively.

The difficulty of estimation of these models grows with the number of discrete choices, so dedicated commercial software packages are recommended for their estimation. In our multivariate analysis, MNL and MPL models were implemented via the software package Stata. We refer the reader to Stata base manual [45] for a detailed discussion of the estimation procedure.

Note that during the analysis, five modes of transports are used as the dependent variable (see Figure 3), and two groups of individual level variables: weather and context information attributes are used as the independent (observable) variables (see Table 2). However, in the model, we exclude the use of relative humidity, precipitation, and holidays information because, as seen previously, they have a limited impact on the transport mode decisions. The categories of the remaining weather conditions, such as temperature, wind speed, and air quality, are adjusted according to relationships and patterns observed during the statistical analysis. In fact, we classify the temperature into high (highs greater than 25 °C), warm (highs around 25 °C; lows around 15 °C), mild (highs around 15 °C; lows greater 0 °C), and cold (highs around 0; lows less than 0 °C).

4.2. Results

Table 3. MNL and MNP model results of the impact of weather and context data on transport mode choices.

Observable Variables	Multinomial Logit (MNL)				Multinomial Probit (MNP)
	Bike	Car	Bus	Train	Bike	Car	Bus	Train
Intercept	0.4154	−2.560 **	−1.265 **	−2.928 **	0.320	−1.607 **	−0.788 *	−1.701 **
Temperature (°C)
<0 (cold)	ref	ref	ref	ref	ref	ref	ref	ref
0–15 (mild)	−1.122 **	−1.068	−2.292 **	−1.349 *	−0.974 **	−0.823 *	−1.685 **	−0.981 *
15–25 (warm)	−1.628 **	−1.371 *	−2.152 **	−2.068 **	−1.335 **	−0.998 *	−1.642 **	−1.464 **
>25 (high)	−1.969 **	−1.558 **	−1.331 *	−1.338 *	−1.585 **	−1.144 *	−1.103 **	−1.013 *
Wind speed (m/s)
0–3	ref	ref	ref	ref	ref	ref	ref	ref
3–7.8	0.013	0.311	0.256	0.554 *	0.020	0.177	0.160	0.320 **
>7.8	0.280	0.621	−0.018	0.228	0.200	0.400	−0.004	0.069
Air quality (μg/m3)
<300	ref	ref	ref	ref	ref	ref	ref	ref
>300	0.921 **	−1.333	−0.564	0.405	0.735 *	−0.524	-0.270	0.315
Trip distance (km)
0–4	ref	ref	ref	ref	ref	ref	ref	ref
4–8	2.143 **	2.875 **	2.588 **	2.829 **	1.741 **	1.964 **	1.890 **	1.890 **
8–12	1.912 **	4.222 **	4.305 **	4.926 **	1.480 **	2.829 **	3.082 **	3.150 **
12–20	0.872 *	5.211 **	4.319 **	6.381 **	0.838 **	3.444 **	2.992 **	4.091 **
>20	3.609 **	9.380 **	7.094 **	7.912 **	2.150 **	5.862 **	4.293 **	4.553 **
Olympics (ref. non Olympics)	−0.471	−2.347 **	0.2617	−0.432	−0.380	−1.635 **	0.210	−0.353
Rush hour (ref. non rush hour)	0.022	−0.999 **	−0.259	−1.351 **	−0.005	−0.624 **	−0.146	−0.827 **
Day (ref. night)	−0.140	0.355	0.515 **	0.059	0.102	0.211	0.330 *	0.0382
Modelling performance
Log likelihood	−2498.486				−2505.669
AIC	5108.971				5123.339
BIC	5438.823				5453.19
McFadden R-squared	0.3289				0.012
	** Significant at α < 0.01				* Significant at α < 0.05

shows the results of the both models with a summary of the relationships between the transport mode choices, weather, and the context information. The modeling performance shows that the MNL model performs better than MNP because it has a larger McFadden R-squared value and smaller AIC and BIC values, which shows that the results will be in favor of the MNL model. The values in the tables represent the standardized regression coefficients, with the star indicating their respective statistical significance. Note that the coefficients are a relative measure of transport choice compared to walking, which was the reference mode. For instance, our findings indicate that a negative temperature effect on the bike, car, bus, and train makes them less likely to be used as transport modes than walking.

However, this means that other than the sign, the coefficients do not have a lot of useful interpretation since the magnitude of the coefficients cannot be interpreted or quantified. Thus, the marginal effects of the observatory variables $z_n$ on the dependent variable i are required. Marginal effects are defined in (6) as the amount of change in dependent variable due to one unit change of the observatory variable in the model system.

$$\begin{array}{c}\hfill \frac{\partial P_{in}}{\partial z_n}=P_{in}({\gamma }_i-{\overline{\gamma }}_i)\end{array}$$

(6)

Table 4. Marginal effects of MNL and MNP models of the impact of weather and context data on transport mode choices.

Observable Variables	Multinomial Logit (MNL)					Multinomial Probit (MNP)
	Walk	Bike	Car	Bus	Train	Walk	Bike	Car	Bus	Train
Temperature (°C)
<0 (cold)	ref	ref	ref	ref	ref	ref	ref	ref	ref	ref
0–15 (mild)	0.234 **	−0.006	0.003	−0.219 **	−0.013	0.273 **	−0.050	0.077	−0.218 **	−0.011
15–25 (warm)	0.320 **	−0.107	−0.004	−0.177 **	−0.031	0.351 **	−0.135	0.004	−0.183 **	−0.037
>25 (high)	0.279 **	−0.216 **	−0.026	−0.030	−0.006	0.319 **	−0.234 **	−0.025	−0.053	−0.005
Wind speed (m/s)
0–3	ref	ref	ref	ref	ref	ref	ref	ref	ref	ref
3–7.8	−0.039	−0.024	0.018	0.027	0.018 *	−0.037	−0.018	0.0158	0.020	0.020 *
>7.8	−0.062	0.039	0.046	−0.025	0.002	−0.057	0.0342	0.05	−0.024	−0.002
Air quality (μg/m3)
<300	ref	ref	ref	ref	ref	ref	ref	ref	ref	ref
>300	−0.099	0.312 **	−0.125 **	−0.096 *	0.005	−0.100	0.295 **	−0.121 **	−0.084 *	0.010
Trip distance (km)
0–4	ref	ref	ref	ref	ref	ref	ref	ref	ref	ref
4–8	−0.517 **	0.303 **	0.061 **	0.125 **	0.027 **	−0.518 **	0.298 **	0.062 **	0.128 **	0.028 **
8–12	−0.629 **	−0.015	0.123 **	0.398 **	0.123 **	−0.629 **	−0.017	0.125 **	0.394 **	0.126 **
12–20	−0.664 **	−0.162 **	0.225 **	0.246 **	0.353 **	−0.662 **	−0.163 **	0.229 **	0.245 **	0.350 **
>20	−0.725 **	−0.176 **	0.672 **	0.162 **	0.066 **	−0.724 **	−0.177 **	0.671 **	0.163 **	0.067 **
Olympics (ref. non Olympics)	0.130 *	−0.046	−0.175 **	0.100 **	−0.004	0.140 *	−0.040	−0.21 **	0.115 **	−0.005
Rush hour (ref. non rush hour)	0.072	0.058 **	−0.071 *	−0.014	−0.044 **	0.063	0.0503 **	−0.073 **	0.004	−0.052 *
Day (ref. night)	−0.031	−0.064**	0.024	0.073*	−0.006	−0.029	−0.057 *	0.023	0.064 *	0.010
	** Significant at α < 0.01					* Significant at α < 0.05

reports the marginal effects for each variable according to the MNP and MNL models. The positive values represent an increase in the probability of selecting an alternative by the marginal effect expressed as a percentage, while the negatives indicate the opposite. It should be noted that qualitative results are consistent between the MNP and MNL.

5. Discussion

In the previous section, we presented statistical models that help in understanding the relationship between the weather and context information with transport mode choices. In this section, the results in Section 4.2 are discussed.

5.1. Air Quality

Although the air quality in Beijing has been documented to be very poor because of excessive emissions from vehicles and industries [46], there is limited literature on how it influences the transport mode decisions. In fact, the only work available by Mar et al. [17] studies the effect of air quality on transport mode choices, but for middle school students. This fact motivated this present study.

Our study shows that when the air quality changes from the very unhealthy category to hazardous, this change will increase the bike share by 31% and reduce the car and bus shares by 12% and 10%, respectively. The likely explanation is that hazardous air quality causes serious health concerns, such as loss of lung capacity and decreased lung function, according to epidemiological studies. Therefore, the government in Beijing introduces vehicle controls to reduce CO₂ emissions to prevent the population from long-term exposure to polluted air. Consequently, the population is encouraged to use environmentally friendly transport mode choices, such as the bike, to get to their preferred destinations.

However, previous research of Mar et al. [17] asserted that when the air quality is hazardous, a decrease in the bike and walk shares and an increase in the public transport and car shares for middle school students are observed. The participants in the Geolife dataset are assumed to have been adults, so we see that the results in our study are different from the results in [17] because children are adversely affected by very poor air quality, compared to adults [47]. The increase in car and public transport shares and the subsequent decrease in the bike and walk shares are meant to prevent children from direct exposure to air pollution.

5.2. Temperature

The results of our study show that when the temperature increases, people walk more, which decreases the relative likeliness of choosing other transport modes. Table 4 shows that with a unit increase for mild and warm temperatures, the probabilities of walking is expected to increase by 23% and 32%, respectively, while the probabilities of choosing the bus is expected to decrease by 21% and 17%, respectively. This result is in line with previous studies, which indicate the positive effects of temperature on walking or biking shares [21,48,49]. However, a unit increase for high temperatures would reduce the bike share by 21%, while the walking share would increase by 27%, which in line with similar studies made in Rotterdam [49], Washington DC [50], Brisbane [51].

Interestingly, some studies add that temperature may either have a bell-shape or rather a linear shape influence on particular transport mode choices [38,49,51]. To detect this possibility, we modeled the temperature effects on the walk, bike, and bus transport mode choices in five degree steps (see Figure 6). Our findings indicate that the relationship between temperature and the changes in the transport modes for walking, biking, and the bus is linear. Moreover, we see that temperature has a positive linear relationship with walking and taking the bus, and a negative linear relationship with biking.

5.3. Wind Speed

Table 4 shows that a unit increase in the wind speed between 3 and 7.8 m/s would result into an increase of 1.8% in the train modal share. This result shows that the wind speed, just like relative humidity and precipitation, has a very limited effect on transport mode choices in Beijing. The possible reason is that during the period in which this study was conducted, the average wind speed in Beijing was very light, according to the Beaufort scale (see Table 2).

However, our results are in contrast to previous research, which established a statistically significant relationship between wind speed and several transport mode choices [17,52,53]. The results of these studies mostly state that walk and bike shares decrease with increasing wind speed. For instance, the authors in Mar et al. [17] showed that with an increase in wind speed, the probability of walking decreases, while the probability of using a car increases for school-going children in Beijing.

5.4. Trip Distance

The trip distance is known to influence transport mode choices. Generally, literature finds that built environment can be a mediator variable between trip distances and transport mode choices. The reason is that the distances to the destinations relative to residential areas largely depend on the land use and density of the built environment [54,55].

However, the authors in [13,17] directly model the impact of trip distances on transport mode choices without considering the land use classification or the socio-economic attributes as mediators between trip distances and transport mode choices. In fact, Mar et al. [17] goes a step further to suggest that there are no fixed rules on modelling the relationship between transport mode choices with an increase in trip distance.

Given that the Geolife trajectory dataset is anonymized, we directly model the influence of trip distances on transport modes in line with the analysis in [13,17]. However, future works must integrate GPS trajectory data with other data sources or capture more details of the users, such as demographics, individual characteristics, household characteristics, and trip characteristics, to improve the performance of the models.

Looking at the results from the direct influence of trip distances on transport mode choices in Figure 7, we see that trip distances significantly impact all transport choices. For instance, a unit increase in distances of 4–8 km is expected to decease walking by 51%, while increasing the bike, bus, car, and train shares by 30%, 6%, 12.5%, and 2.7%, respectively.

However, a unit increase in distances of 8–12 km is expected to lead to a further decrease in walking by 62%, while increasing car, bus, and train shares by 12.3%, 39.8%, and 12.3%, respectively. For a unit increase in distances of 12–20 km, walking and biking are expected to decrease by 66% and 16.2%, respectively, while car, bus, and train shares are expected to increase by 22.5%, 24.6%, 35.3%, respectively.

Beyond 20 km, this trend is similar, except that there is a further increase in the car, bus, and train shares by 67.2%, 16.2%, and 6.6%, respectively, while a further reduction in walking and bike shares by 72.5% and 17.6%, respectively, is expected. These results are in line with the findings in survey studies by [13,17], whose authors found that shorter distances are likely to be performed by walking and bike, while the longer distances by car, bus, and trains.

5.5. Olympics

Generally, it is highly likely that different transport modes will be disrupted due to the magnitude of people that such an event as the summer Olympics attracts. Given that there is no study in the literature that analyzes how a onetime intervention affects the transport mode choices during constantly present factors, such as the weather, this motivates our work.

Our results show that during the Olympics, there is a decrease of 17.5% in the car share and an increase of 13% and 10% in the walk and bus shares, respectively. The decrease in the car share can be attributed to the private car restrictions imposed by the government, and the increase in the bus share is explained by the additional bus lines put in place as an alternative to accommodate the number of passengers. Our findings are consistent with the findings of the authors in [35], who mentioned that during the preparation for the 2008 summer Olympics, a series of traffic restrictions on private vehicles were introduced by the Beijing government.

5.6. Rush Hour

Beijing is known to be heavily gridlocked during rush hour. The government of Beijing tries to solve the traffic congestion through the restricting of new car license plates and the banning of cars from outside Beijing during rush hour.

Our results show a reduction in the likelihood of traveling by car and train by 7.1% and 4.4%, respectively, and an increase in biking by 5.8%. Our findings are consistent with the reduction in car use due to car restrictions imposed during rush hour [56,57].

However, a more general explanation for the reduction in the use of the train may have to do with the need to reduce or avoid passenger congestion in trains or subways. The work by Huang et al. [58] mentions that passenger congestion is a huge challenge during peak hours because millions of people use the trains as one of their primary transport modes in Beijing.

5.7. Day/Night

During the day, Table 4 shows that the probability of biking decreases by 6.4% and the use of the bus increases by 7.3%. A possible reason is that during the day, there are more buses operating than in the night. During the night, there is an increase in the people biking because transport operators reduce the number of buses.

However, our result is contrary to the findings obtained using survey data in Scandinavian cities of Stockholm and Oslo [13], in which during the night, there are fewer trips by active transport modes, such as biking than during the day.

6. Conclusions

Our main objective is to analyze the potential of using GPS trajectories data as an alternative to travel surveys for studying the impact of weather and context data on transport mode choices. In the methodology, we link GPS trajectories with weather and context information and then analyze their relationship by descriptive means and statistical models. Our conclusion are as follows.

First, the analysis of GPS trajectories together with weather and context data provides a means to quantitatively model the choice of transport mode. Our approach does not rely on subjective data, such as the those collected through travel surveys.

Second, regarding context data, the trip distance is the factor that has the greatest influence in the choice of transport mode, while temperature has the most prominent effect of among the weather conditions.

Lastly, policies on traffic restrictions and large events, e.g., the Olympics, affect the choice of transport mode, which can be measurable through GPS trajectories. Our work will support transport mode operators and policy makers in understanding the effect of their design choices and better their planning to suit passengers’ mobility needs based on weather and context data.