THE IMPACT OF WEATHER ON BICYCLE RISK EXPOSURE

Traffic volume is the main independent variable of risk exposure in road safety models. Cyclists as a vulnerable road users are more exposed to weather conditions than e.g. car drivers. As a result, their decision of whether to cycle is strongly related to weather conditions. It suggests that any change in the weather may have a significant effect on bicycle use. Objective of the paper was to indicate which weather parameters have a significant impact on bicycle use, how a change in weather parameters affects the change in bicycle volume (risk exposure) and, consequently, predicted number of crashes with cyclists and which factors differentiate the impact of weather conditions on bicycle volume. The impact of weather on bicycle volume variability was estimated based on literature review. The Web of Science, Scopus and TRID databases were searched. Finally, 33 papers from 1977 up to 2020, different in terms of the methodology used, country of origin, and analyzed group of cyclists, were reviewed. The impact of change in weather conditions on the predicted number of crashes with cyclists was estimated using own road safety models and previous research results. Results indicate that air temperature, precipitation, sunshine, cloud cover, humidity, and wind strength, have a significant influence on bicycle use. The impact of the weather on bicycle volume differs between different cyclists’ groups (different levels of experience, age, gender), trip motivations (recreational, commuting, etc.) and locations (countries, cities, climate zones). The paper shows complexity of impact of weather conditions on cycling and sensitivity of relationship between weather conditions and bicycle volume (i.e. risk exposure) and, as a consequence, bicycle safety. Results indicate that weather conditions should always be taken into consideration when analyzing cycling, especially in road safety analysis. The discussion of presented research results, research methods used with their limitations, and recommendations for future research were described.


Introduction
The bicycle is a healthy, low-cost and environmentfriendly mode of transportation. Individual features (e.g. income, gender, age), initiatives and policy of local authorities, presence and type of bicycle infrastructure, perceived risk of injury, and the presence of bikeshare systems are some of the many factors affecting bicycle volume. One of them is also the weather. Cycling is more sensitive to weather conditions compared with other modes of transportation (Sabir, 2011), (Miranda-Moreno and Nosal, 2011). More than half of cyclists (58%) consider the weather when deciding whether to bike (Gallop, Tse and Zhao, 2012). The daily fluctuation of bicycle volume in 80% is described by weather conditions (Thomas, Jaarsma and Tutert, 2013). It suggests that any change in the weather may have a significant effect on bicycle use. Traffic volume is one of the main independent variables in road safety analysis (Gaca, 2002), (Li et al., 2016). Research on impact of weather conditions on number of bicycle crashes or injuries was conducted previously (Kim et al., 2007), (Klop and Khattak, 1999), (Prati et al., 2017). However, those research did not include a change in cyclists' risk exposure in different weather conditions. That approach would give a more insightful estimation of the impact of weather on cyclists road safety. Aside from road safety analysis, bicycle volume data is required in infrastructure planning and designing (estimations of bicycle traffic distribution, calculations of traffic performance and traffic signals program, etc.). When planning or designing road infrastructure, traffic volume has to represent traffic conditions in a long-term period. Therefore, change in bicycle volume as a result of climate change should be taken into consideration. Additionally, estimation of bicycle volume variation due to change in weather conditions enables for a more appropriate comparison of bicycle volumes in different locations. It is necessary when the increase in cyclists volume is estimated. Therefore, knowledge of impact of weather conditions on bicycle volume is necessary in various bicycle traffic analysis. The aim of the paper was to indicate which weather parameters have a significant impact on bicycle volume and how change in weather parameters can affect bicycle use (i.e. bicycle risk exposure) and, consequently, predicted number of crashes with cyclists. The impact of weather on bicycle volume variability was estimated based on literature review. The Web of Science, Scopus and TRID databases were searched. Finally, 30 papers from 1977 up to 2019, different in terms of the methodology used, country of origin, and analyzed group of cyclists, were in detail reviewed. The discussion of presented research results, research methods used to evaluate impact of weather on bicycle volume and their limitations were described. The impact of change in weather conditions on risk exposure and therefore predicted number of accidents involving cyclists was estimated using own road safety models and previous research results. The paper also includes recommendations for future research. Presented results show complexity of impact of weather conditions on bicycle volume and sensitivity of this relationship. The paper is informative for road administration, designers and transport planners.

Impact of weather conditions on bicycle use 2.1. Literature review method
The impact of weather on bicycle volume variability was estimated based on literature review. A few databases were searched to gather research papers, i.e. Web of Science (WoS), Scopus and Transport Research International Documentation (TRID). The keywords used in searching were: "bicycle", "bicycle volume" combined with "weather" or "environment". Table 1 presents the number of hits for each used search term in different databases. The total number of found papers was 1150. conditions on bicycle volume? Due to different climate and transportation culture, impact of weather on cyclists volume may differ between different geographic areas (countries, cities) (Ashqar, Elhenawy and Rakha, 2019). Therefore, papers with a different country of origin were included. Finally, 33 papers from 1977 up to 2020 were reviewed in detail and described in the paper. They varied in terms of the methodology used (different methods of bicycle volume data gathering and analysis) and analyzed groups of cyclists (different level of experience, age, gender, etc.). Figure 1 shows the distribution of reviewed papers published in successive years.

Summary and discussion of previous research results
Research results on impact of weather conditions on bicycle use are presented in details separately for survey (Table 2) and empirical research (Table 3) of bicycle volume gathering. In tables detailed information about source of data, year of publication, country of origin, method of bicycle volume gathering and data analysis, and research results are presented. For some papers in Table 2 additional notes are provided, which help understand how impact of weather on cyclists was evaluated. In subsubsections from 2.2.1 to 2.2.5 summary and discussion of presented research results are provided.  cyclists were asked to rate several factors (e.g. risk of injury in car-bike collision, possibility to make the trip in daylight hours, beautiful scenery of the route) in terms of its influence on their likely to cycle; score -1 meant much less likely to cycle, score -0.5less likely to cycle; score 0 no influence on the decision to cycle; score 0.5more likely to cycle, score 1much more likely to cycle total score for hot and humid weather -0. 16 itation, errands, darkness, car park, environment, and exercise) for mode choice when travelling to work by bicycle; score 1 denoted "no importance" and 7 denoted "great importance" impact of weather conditions differs between cyclists with different experience (winter cyclists, summer-only cyclists and never cyclists); in a questionnaire survey in 1998, temperature had the biggest impact on summer-only cyclists (score 5.47) and it was the most important factor; the score for temperature was 3.22 (8th place out of 11) for winter cyclists. and 3.29 (5th place out of 11) for never cyclists; in 1998 precipitation was ranked in 7th place by winter cyclists (score 3.34), in 2nd place by summer-only cyclists (score 5.27), and in 4th place by never cyclists (score 3.53); in 2000 scores for temperature and precipitation were similar to those obtained in 1998; temperature and precipitation were scored 2.54 (9th place out of 10) and 3.1 (7th place), 5.16 (3rd place) and 5.71 (1st place), 3.78 (4th place) and 4.26 (3rd place) by winter cyclist, summer-only cyclist and never cyclists, respectively; total score for temperature and precipitation (for all cyclists) was 3.91 (4th place out of 11) and 4.0 (2nd place), 3.59 (7th place out of 10) and 4.16 (5th place) in the surveys conducted in 1998 and 2000, respectively; darkness was the least important in mode choice decision; total score for darkness was 2.47 and 2.37 in the survey in 1998 and 2000, respectively; darkness had the strongest impact on summer-only cyclists and the weakest on winter cyclists  forecasted daily high air temperature had a positive effect and forecasted clouds, rain, snow, thunderstorms, or weather warnings had a negative effect on bicycle volume; actual weather and forecasted weather are both important determinants of bicycle use; bicycle volume was lower by 3,6% and higher by 11,5% for hours with no rain but for which rain was forcasted and for hours with rain but for which rain was not forcasted, respectively, compared to hours with correctly forecasted rain; even if midday and afternoon hours were predicted to be rain-free, rain forecasted for the morning hours significantly reduced bicycle volume ; impact of weather (actual and forecasted) for recreational counting stations was stronger than for utilitarian ones.

Weather parameters affecting bicycle use
Weather parameters found to have a significant impact on cyclists' volume are the following: air temperature, precipitation, sunshine, cloud cover, humidity, and wind strength. Impact of weather conditions on bicycle use was previously described in reference to daily number of trips made by public bicycle (Noland  and Ishaque,  could be the result of lack of extreme temperatures in analysis period. Air temperature, found to have major impact on bicycle volume in many research, was not statistically significant in (Corcoran et al., 2014). The reason might be small variability of temperature or the way it was analyzed (i.e. two categories of temperature were taken into consideration: low (up to 21 o C) and high (above 21 o C)). Variability of weather parameters is also an effect of adopted analysis period. For example in (Hanson and Hanson, 1977) 39 days of the analysis was too short to find strong relationships between cycling and weather conditions. In (Sabir, 2011), (Nankervis, 1999), (Bergström and Magnusson, 2003), (Lewin, 2011) variability of bicycle volume in terms of season change (spring, summer, autumn, winter) was analyzed. It is worth to mention that season is strictly correlated with air temperature, precipitation, and humidity. Using seasons rather than weather parameters is an indirect method to evaluate their impact on cycling. However, correlation coefficients between seasonal and weather parameters should be analyzed. Using the seasonal coefficient of variation of bicycle volume enables a general estimation of change in bicycle use due to changing weather conditions (seasons). Nevertheless, it is not useful when changes in bicycle volume are analyzed in a shorter period of time (week, day, hour).

Methods of bicycle use data gathering
In the previous research two main methods of bicycle use data gathering were used, i.e. survey and empirical data (from automatic counters or bikeshare systems). Survey research were done in two different ways: − survey of the preferences (Ahmed, Rose and Jacob, 2013), (Winters et al., 2006), (Winters et al., 2011), (Bergström and Magnusson, 2003)show general trends and tendencies to make a trip by bicycle in various weather conditions, allow to assess and rank weather factors in terms of its impact on cycling in comparison to other factors affecting mode choice (infrastructure, time of travel, etc.), often made including different travel motivations, age groups, gender and cycling experience. In these studies, the assessment is often carried out in a qualitative way, e.g. in (Winters et al., 2011) and (Bergström and Magnusson, 2003) authors used scores from -1 up to 1 (every 0,5) and from 1 to 7 (every 1) respectively. The results of these studies do not allow to estimate the change in bicycle volume for given values of weather parameters, but give in-depth insight into the decision-making process whether to cycle. Therefore results of those survey are a valuable complement to empirical research and can be used by road administration (what can we do to encourage people to cycling?, on what group of potential cyclists should we focus?); − research based on trips that were actually made (Sabir, 2011), (Richardson, 2000), (Saneinejad, Roorda and Kennedy, 2012)analyzed together with weather parameters data from weather stations. Relationship developed in these type of studies allow to evaluate change in bicycle volume for given values of weather parameters. Not only weather conditions, but also age, gender, physical fitness, travel motivation, accessibility to bicycle, cycling experience, trip distance, type and standard of bicycle infrastructure, perceived safety level, presence of parking, public transport fares, time of trip, etc. have an impact on mode choice. Therefore, when planning survey research a great effort should be put into choosing the right group of respondents (Saneinejad, Roorda and Kennedy, 2012), (Flynn et al., 2012). When assessing the potential increase in bicycle volume, it should be remembered that some people will not choose bicycle as a mode of transport, no matter what actions will be taken (improvement of bicycle infrastructure, etc.). Automatic counters are the main source of bicycle volume data. Cyclists can use dedicated infrastructure, roadway or pedestrian paths. Using bicycle volume data from automatic counters is a adequate approach for separated bicycle paths, but may generate error when bicycle infrastructure located next to pedestrian path or a roadway is analyzed. Another limitation in automatic counters usage is an error arising when two or more cyclists ride together or ride with high speed. To eliminate those problems bicycle GPS data could be used in the analysis (Pogodzinska, Kiec and D'Agostino, 2020).

Methods of weather data gathering
Based on (Gallop, Tse and Zhao, 2012), 58% of cyclists considered the weather when deciding whether to cycling, and 77% of them based their decision on current rather than forecasted or recent weather. Among respondents who based their decision on forecasted weather, 41% checked it just before they leave, 24% up to 2 hours before and 29% on the evening before. For comparison, (Ahmed, Rose and Jacob, 2013) found that about half of the respondents planed which days they will bike ride in advance and 49% of commuters considered current weather conditions as well as forecasted weather when planning their trips. It should be noticed that forecasted weather, which is also the basis of decision making process, can be different than data from weather station, which were mainly used in the previous research. In previous research, models describing impact of weather on bicycle use were developed using both quantitative and qualitative weather measures. In (Gallop, Tse and Zhao, 2012) if fog, rain, snow or drizzle were present, dummy variable was 1 (if not, it was 0). Nankervis used categories of rain, wind and temperature (Nankervis, 1999), Saneinejad et al. distinguished five sky conditions and nine temperature categories (Saneinejad, Roorda and Kennedy, 2012), Brandenburg et al. used two precipitation categories (with and without precipitation) and thermal index (Physiologically Equivalent Temperature, PET) based on temperature categories. Using categories rather than direct measures of weather variable may impede finding relationship between that parameter and bicycle volume, like it could be in (Corcoran et al., 2014), where air temperature was not statistically significant determinant of bicycle volume. It can also not allow to observed non-linear impact of weather parameter on bicycle use. Different measures of weather parameters should be taken into consideration. For example in majority of previous research rain was represented by its amount in mm. However, using intensity of the rain (in mm per hour or day) rather than its amount can help find new and different relationships (Sabir, 2011). Interesting approach was used in (Phung and Rose, 2007), (Ahmed, Rose and Jacob, 2010) where apparent temperature (dependent on humidity, wind strength and air temperature) rather than air temperature alone was analyzed.

Methods of data analysis
Research based on observed data were conducted with reference to hourly or daily bicycle counts. In general, compared to models build for daily volumes, relationships for hourly volumes were worse fitted to empirical data (analysis with reference to hourly data: R 2 = 0,38-0,59 (Miranda-Moreno and Nosal, 2011); R 2 = 0,30-0,60 (Ahmed, Rose and Jacob, 2010), R 2 <0,19 (Gebhart and Noland, 2014); analysis with reference to daily data: R 2 = 0,79 (Thomas, Jaarsma and Tutert, 2013), R 2 = 0,52-0,71 (Brandenburg, Matzarakis and Arnberger, 2007), R 2 =0,68 (Mathisen, Annema and Kroesen, 2015), R 2 =0,78-0,85 (Lewin, 2011)). In (Tin et al., 2012) weather factors explained 23% and 56% of the variability of hourly and daily bicycle volume respectively. However, when using daily data impact of change of weather conditions during the day on bicycle use cannot be analyzed. According to (Thomas, Jaarsma and Tutert, 2009), total amount of precipitation for wet night followed by a sunny day and dry night followed by a wet day may be the same, however impact on bicycle use can be different. Therefore, like Richardson suggested (Richardson, 2000), because most bicycle trips are made in daylight hours, it may be better to use daylight-hour rainfall in the analysis. The same could be adopted for other weather parameters. In previous research, regression models were the main method of data analysis. In (Gallop, Tse and Zhao, 2012) two methods were implemented i.e. linear regression and ARIMA model, characterized by R 2 =0,35 and R 2 =0,95, respectively. It suggests that using more sophisticated methods of data analysis can help to develop model better fitted to empirical data.

Factors differentiate the impact of weather on bicycle volume
Cyclists are more exposed to weather conditions than e.g. car drivers, and therefore their decision of whether to cycle is strongly related to personal comfort. As mentioned in (Sabir, 2011), (Thomas, Jaarsma and Tutert, 2013), (Brandenburg, Matzarakis and Arnberger, 2007), (Saneinejad, Roorda and Kennedy, 2012), (Hanson and Hanson, 1977), (Bergström and Magnusson, 2003), (Thomas, Jaarsma and Tutert, 2009), (Gebhart and Noland, 2014) personal comfort, and as a result the impact of weather conditions on bicycle volume, differ between cyclists' groups (different level of experience, age, gender) and trip motivation. In general, recreational cyclists (who are less experienced and ride from time to time) are more sensitive to bad weather conditions than commuting cyclists (who are more experienced and ride frequently). It may be confusing that some of the research results described in the paper show such a different impact of weather on bicycle volume, even though they were made in the same country or even city. For example, in (Richardson, 2000) author calculated that daily rainfall of around 8 mm resulted in a 50% decrease in bicycle volume, compared to days with no rain. On the other hand, in (Phung and Rose, 2007) it was found that cyclists' volume decreased by 8-19% if daily rainfall was 0.2-10mm. Both research was conducted in the city of Melbourne (Australia). Nevertheless, research (Richardson, 2000) was based on a questionnaire survey made in 1994 and research (Phung and Rose, 2007) was made over 10 years later, based on a data from automatic counters. Moreover, in (Phung and Rose, 2007) wind with strength of 40-62kph resulted in a reduction in commuter cyclist volume by 11-23%. On the other hand, (Corcoran et al., 2014) showed that wind with strength already above 5 km/h reduced the number of bicycle trips by 17%. Research by (Corcoran et al., 2014) was also conducted in Australia, however in the city of Brisbane and in reference to bikeshare system users. Different methodology used and time of data gathering may be the explanation for the observed differences in research results. It shows that implementation of models describing the relationship between weather conditions and bicycle volume developed for different locations and time, related to specific group of cyclists should be done very carefully.

Evaluation of impact of weather on pre-
dicted number of crashes with cyclists Review of previous research results indicates that bicycle volume is significantly influenced by weather conditions. The impact of weather on a predicted number of accidents involving cyclists was estimated using own road safety models (not published) and previous research results.

Evaluation based on own safety models
Based on inventory of over 50km of selected street sections in City of Cracow, a database of factors that may affect cyclists' safety for 171 homogenous road segments was collected. For each homogenous segment, the database included: road parameters (length of the segment, street function, the number of lanes and roadways, speed limit), type and standard of bicycle infrastructure (width, type of the pavement, offset from the roadway edge, bicycle traffic separated or mixed with other road users), access to road (the number of public and residential access points), crossings with pedestrian and vehicles (the number of crossings, traffic management at crossings), public transport (the number of bus stops), parking (presence and angle of parking, the number of parking spots), as well as number of crashes with cyclists (based on crash data from 2015-2017) and Annual Average Daily Bicycle Traffic (AADBT) estimated based on bikeshare system GPS data (Pogodzinska, Kiec and D'Agostino, 2020). Three Generalized Linear Models (GLMs) with negative binomial distribution of dependent variable i.e. number of crashes with cyclists, were developed. The models forms are shown in equations 1-3 and the results of calibration are shown in Table 4.
AADBT and length of road segment are independent variables in all models. Additionally, to evaluate impact of various infrastructure characteristics on cyclists safety, each model includes different independent variable i.e. cross section-Model 1, offset of bicycle traffic from the roadway edge -Model 2, type of bicycle infrastructure -Model 3. Figure 2 shows a relative change of predicted number of crashes with cyclists due to relative change of AADBT. It should be mentioned that AADBT is also indirectly included in independent variables i.e. crosssection, type of infrastructure used by cyclists and its offset from a roadway edge. Therefore, models present differences in impact of AADBT change on predicted number of crashes with cyclists. However, those differences are not significant. Model 3 is the least and Model 1 is the most sensitive on AADBT change. Relative change of AADBT in range 0,5-2,0 100 Pazdan, S., Archives of Transport, 56 (4), [89][90][91][92][93][94][95][96][97][98][99][100][101][102][103][104][105]2020 results in relative change of predicted number of crashes with cyclists in range 0,72-1,38. For example, if AADBT increase by 10%, predicted number of crashes with cyclists increase by about 5%.
Presented models can be used to evaluate impact of weather in a longer period (e.g. impact of rainy summer or warmer winter, which result in change in AADBT) or when climate change is considered.

Evaluation based on previous research results
To evaluate impact of weather on bicycle use in a shorter period bicycle crash models described in (Kröyer, 2016), (Schepers et al., 2011), (Amoh-Gyimah, Saberi and Sarvi, 2016) were used. Table 5 presents example changes in predicted number of crashes with cyclists due to changes in weather conditions. For example, based on models from (Tin et al., 2012) and (Kröyer, 2016), if temperature increases by 5 o C, daily bicycle volume increases by 13% (2,6%*5), and therefore a number of bicycle single crashes increases by 9%. For comparison, according to models developed in (Flynn et al., 2012) and (Amoh-Gyimah, Saberi and Sarvi, 2016), if temperature increase by 5 o C likelihood of commuting to work by bicycle increase by 15% (3%*5), and therefore number of crashes with cyclists increase by 6%. Presented calculations show that small change in weather conditions, especially in air temperature and precipitation, results in significant change in predicted number of crashes with cyclists.