Главная Обзор Помощь
Вход
ardo
/
Obsidian
1
0
Ответвить 0
Файлы Задачи 0 Запросы на слияние 0 Вики
Дерево: a0e0ef701f
Ветки Метки
Obsidian
/
Zotero
/
001_artiklid
/
Road Safety Indicators
/
paganyspatiotemporalriskpredictionwildlife2022.md

paganyspatiotemporalriskpredictionwildlife2022.md 94 KB
История Исходник

category: literaturenote citekey: paganyspatiotemporalriskpredictionwildlife2022 title: A spatiotemporal risk prediction of wildlife-vehicle collisions using machine learning for dynamic warnings authors: "Pagany, Raphaela" year: 2022 date: 2022-00-00 2022 doi: 10.1016/j.jsr.2022.09.001 publication: Journal of Safety Research url: "https://doi.org/10.1016/j.jsr.2022.09.001" zotero_key: 3X2N5HHA zotero_storage: WFJTHIB6 collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Pagany, Raphaela"

status: converted

Journal of Safety Research

journal homepage: www.el sevier.com/locate/jsr

A spatiotemporal risk prediction of wildlife-vehicle collisions using machine learning for dynamic warnings

Raphaela Pagany

Institute for Applied Informatics, Deggendorf Institute of Technology, Freyung, Germany Interfaculty Department of Geoinformatics, University of Salzburg, Salzburg, Austria

article info

Article history: Received 12 September 2021 Received in revised form 13 May 2022 Accepted 1 September 2022 Available online 16 September 2022

Keywords: Wildlife-vehicle collision Machine learning Spatiotemporal prediction Random forest Neural networks Support vector machine classifier

abstract

Introduction: The technology in the automotive industry is becoming increasingly safer in the age of automated driving, but the number of accidents is still high, especially in wildlife-vehicle collisions (WVCs). To better avoid these accidents, patterns involved in these accidents must be detected. Method: This paper presents a spatiotemporal risk prediction of WVCs, including various road and environmental characteristics. A process of data preparation using GIS automated by Python scripts was developed to enable a spatiotemporal link of diverse features for the subsequent predictive data analysis. Different machine learning (ML) approaches were applied- random forest (RF), feedforward neural networks (FNN), and support vector machine classifier (SVM) - including automated ML to predict the risk of WVCs. Therefore, a dataset of approximately 731,000 accidents reported to the police in Bavaria over a period of 10 years (2010–2019) was used. In addition, non-accidents were randomly generated in Python over time and space for the supervised ML processes. As the actual risk probability for WVCs and non-WVCs is not entirely known, the impact of different training ratios between accidents and nonaccidents was tested on the risk prediction quality (RPQ) (25%, 50%, 75%, 90% WVCs) of the doubleweighted sensitivity and single-weighted specificity rate. Results: The test yielded high mean values of RPQ as an indicator for a suitable WVC prediction. Both RF (86.6%) and FNN (86.7%) were identified as suitable choices for WVC risk prediction in terms of RPQ. The SVM yielded a lower prediction quality, even though acceptable results could be achieved within a shorter runtime. Practical Applications: Spatial transferability was verified since the algorithm was trained on the dataset of Bavaria (excluding Upper Bavaria) and successfully tested in Upper Bavaria. WVC forecasts were also proven through training with datasets from 2010-2017 and in prediction for 2018 and 2019.

2022 National Safety Council and Elsevier Ltd. All rights reserved.

1. Introduction

Injuries and fatalities due to traffic accidents with wildlife have a great impact on the society and wildlife ecology. Vehicle damages cause significant financial costs, even when the actual numbers of wildlife-vehicle collisions (WVCs) are often unknown. German vehicle insurers registered nearly 300,000 accidents with wildlife in 2019 - more than ever before (GDV, 2020). In total, WVCs cost German insurers around 885 million euros. In other countries, the WVC rates are also estimated to be significantly high, for instance, four million in Belgium (Morelle, Lehaire, & Lejeune, 2013), 1.5 million collisions with deer in the United States (Sullivan, 2011), and five million with amphibians and reptiles in Australia per year (Duellman, 2001). But while the collisions are mostly counted in small regions, nationwide or international numbers are estimated to be several times higher than the reported collision numbers (Gkritza, Baird, & Hans, 2010; Steiner, Leisch, & Hackländer, 2014).

Predictive analysis seems to be a useful approach to analyze WVC risk and its influential conditions in a spatiotemporal manner aiming to transfer the risk prediction model globally, for superregional mitigation possibilities. Subsequently, the transferability of such a model would decimate the need for WVC data as a prerequisite for adequate measures in specific areas. By studying the environmental characteristics of the roadway, topography or weather, patterns can be identified to determine if there is a risk of an accident. By using these patterns, predictive analyses enlarge the knowledge about the accident risk, even if little is known about the accident or non-accident locations.

This study applies three ML algorithms - random forest, feedforward neural networks, and support vector machine classifier - to achieve a spatiotemporal WVC risk prediction model and, hence,

E-mail address: raphaela.pagany@th-deg.de

a basis for greater road safety and wildlife conservation. By using GIS tooling automated with Python, heterogeneously structured data regarding the time and the environmental conditions can be brought to one key reference. In this feature engineering process, different spatiotemporal resolutions were harmonized to obtain a pre-processed dataset for further analyzing steps. Road, infrastructure, land use, weather, topographical and further geo-data were identified as potentially influential. A spatiotemporal resolution of one hour and approximately 25-metres road sections was chosen for a detailed WVC analysis. The predictive quality of the algorithms, and their spatial and temporal transferability were examined. For an improved prediction, automated ML (AutoML) was applied. This study addresses three research questions as follows: (1) Are ML methods suitable for an appropriate risk prediction WVCs? (2) How must the training data be structured to enable a high quality of WVC risk classification? (3) Can the approach be transferred temporally and spatially?

The paper is organized as follows: After the literature review (Section 2), the input data for the prediction model and the data pre-processing are described, before the case study region is presented (Section 3). Section 4 describes the analyzing structure including the selection of algorithms and parameters, and evaluation metrics. In Section 5, the results are presented, while in Section 6, the results and advantages of the findings for ecology and traffic management are proposed. Finally, an outlook for future work is given in Section 7.

2. Literature review

Several studies have been employed to address the problem of low data availability since years, either by monitoring carcasses along the road networks (Grilo, Ferreira, & Revilla, 2015; Lutterschmidt, Weidler, & Schalk, 2019), by voluntary citizen science approaches (Heigl et al., 2016; Keken, Sedoník, Kušta, Andrášik, & Bíl, 2019), or by obliging the driver to report the accident to the police or to the local hunter (Snow et al., 2018). Especially in countries with dense road networks, national parks, or high wildlife populations, diverse research teams collected data to study WVCs and their environment (Arevalo, Honda, Arce-Arias, & Häger, 2017; Cherry et al., 2019; Pagany, 2020).

Over the last few decades, various studies constitute that the risk for WVCs is influenced by the drivers' and animals' behavior due to diverse environmental conditions. The temporal factor determines the daily or annual behavior and, thereby, correlates with land use, weather, or light conditions (Garriga, Franch, Santos, Montori, & Llorente, 2017; Kämmerle et al., 2017; Steiner et al., 2014). These dynamical patterns are completed by statistical patterns which together affect the accidents in a spatiotemporal manner. For instance, road conditions, such as the street width, a gentle topography, and the proximity to forest influence the WVC risk, according to a comprehensive review of WVC studies worldwide (Pagany, 2020). In contrast, traffic volume, the distance to urban areas, or road accompanying infrastructure are not clearly assignable influencing or non-influencing factors.

However, most of the studies were undertaken in small areas or considering road sections of only a few kilometers (Gunson, Mountrakis, & Quackenbush, 2011; van der Ree, Jaeger, van der Grift, & Clevenger, 2011), which means that these studies lack information required to mitigate WVCs on a higher level. In addition, the relationship between space and time in WVC is also only sparsely studied. However, this investigation would be essential for dynamically adapted measures in the event of a high WVC risk at a specific time at a specific location, as Pagany (2020) discussed. The review also shows that most studies only analyze a few influencing parameters.

Until now, only a few studies developed a prediction model for WVCs. For instance, Santos, Mota-Ferreira, Aguiar, and Ascensão (2018) developed a Bayesian hierarchical occupancy model that estimated WVC risk, especially in agricultural, open habitats, and within four-lane road sections. Another example is the study of Visintin, van der Ree, and McCarthy (2016), which predicted WVCs with the main determinants being traffic volume, traffic speed, and species occurrence. Both studies explicitly named the advantage of predictive studies, as the analysis decimate the need for a broad data collection. Besides, classic statistics and geo-statistical analyses are chosen in previous studies. They applied hot spot analyses such as the kernel density distribution, as presented in Bíl, Andrášik, Svoboda, and Sedoník (2016), and calculated the density of accidents along certain road section lengths. As further statistical methods for WVC distribution, Malo, Suárez, and Díez (2004) used a Poisson distribution, Valero, Picos, and Álvarez (2015) applied a nearest-neighbor hierarchical clustering, and Tanner, Leroux, and Saunders (2017) used a generalized linear mixed model. Furthermore, Liu, Nieuwenhuis, and McCullagh (2018) and Seo, Thorne, Choi, Kwon, and Park (2015) applied regression analysis to determine the influence of environmental factors such as the roadside land use or seasonal effects.

Studies about other types of traffic accidents used data mining techniques such as machine learning (ML) for identifying the accident risk. Datasets with georeferenced accident occurrences combined with environmental characteristics are often included in the risk analysis. Yang, Chen, and Brown (1999) investigated the severity of road accident injuries in Alabama, United States. A backpropagation neural network was trained including variables such as light conditions and traffic speed to detect safer driving patterns, which may reduce fatalities and injuries by up to 40%. Komol et al. (2021) chose ML-based classification approaches for modeling injury severity of the vulnerable road users: pedestrian, bicyclist, and motorcyclist using k-nearest neighbor, support vector machine, and random forest. They found that motorcyclists have an especially high crash severity. Another study about injury severity is Chen, Song, and Ma (2019). Using a random parameters bivariate ordered probit model, they showed correlations between two drivers' injuries such as driver age, gender, vehicle, airbag or seat belt use or traffic flow. Abdelwahab and Abdel-Aty (2001) applied multilayer perceptron (MLP) and fuzzy adaptive resonance theory (ART) neural networks and included parameters such as the vehicle types and the sex of the driver to predict the severity of the driver's injury in Florida, United States.

In addition, Deublein, Schubert, Bryan, and de Soto Boria (2015) developed an accident risk model using Bayesian Probabilistic Networks including road-related variables, such as the number of lanes, curvature, and traffic characteristics. With a correct prediction of 86.53% of the road segments with a tolerance of 25%, the model could be applied to predict spatial ''black spot" locations regarding the accidental injuries on the Swiss highway network. Beshah and Hill (2010) used a rule mining approach to classify road traffic accidents using adaptive regression trees for a case study in Ethiopia. The study included diverse data about the drivers' age and driving experience, the vehicle age, the road surface, as well as light and weather conditions. Besides, recent approaches can be mentioned that are characterized by similar pre-processing steps of GIS data like for accident risk prediction (e.g., the prediction of connected vehicle movement at intersections). Trajectory movement labelling for predicting driver movements and for warning at intersections ensure advanced transportation safety using ML-based approaches to implement these technologies (Komol et al., 2021; Shrivastava, Verma, & Jain, 2021).

Other studies compared ML techniques to test their quality or predictive accuracy. Chang and Chen (2005), for instance, compared a classification and regression tree (CART) with a negative binomial regression model to analyze the relationship between traffic accidents and highway geometric variables, traffic characteristics, and environmental factors in Taiwan. They found that the CART model is an appropriate method for analyzing the frequency of freeway accidents. Precipitation variables and daily traffic volume are the key determinants for the accident risk. Chong, Abraham, and Paprzycki (2005) compared the performance of neural networks, support vector machines, decision trees, and a concurrent hybrid model involving decision trees and neural networks. The severity of an injury resulting from a traffic accident was predicted. The hybrid approach of decision tree and neural networks outperformed the individual approaches.

The applications and the comparisons of diverse ML techniques to model traffic accidents show that these approaches are suitable for accident risk prediction. Several studies are undertaken to understand the characteristics of driver behavior, road conditions, environmental and weather conditions that are related to the accident probability. However, for a risk prediction, it is not decisive to know the impact of an individual factor, as analyzed by classical statistics, but to develop a decision model considering the environmental factors and the learning characteristics of ML techniques. Thus, ML may also enlarge the knowledge about the accidents with wildlife, as already applied for other road accidents types. A supervised ML classification seems to enable this WVC risk prediction.

3. Material and data preparation

3.1. Data description

For this study, a dataset of 730,867 WVCs that were officially registered by the Police between 2010 and 2019 in Bavaria, Germany was used (StMI & Bavarian State Police, 2020). Each of these registered WVCs from the last 10 years contains a georeference and a timestamp. Additional features were used that are identified in previous studies as influential characteristics for road traffic accidents with wildlife (Pagany, 2020) (Table 1): road-related features, such as road class, road width, infrastructure elements, (e.g., middle barriers, road barriers, and additional lanes), and topographic features (e.g., ditches, or slopes aside the road network). In addition, weather data were used, such as air temperature, precipitation, and cloud coverage, as well as land use data in the prediction model. The land use data are classified in 12 classes as follows: artificial land, cropland seasonal, cropland perennial, forest broadleaved, forest coniferous, forest mixed, shrubland, grassland, bare land, water, wetland, and snow and ice areas. Based on the accident data, time-related features were derived, namely the day of the year, the weekday, the hour of a day, and the solar altitude.

The road information was obtained from the database BAYSIS (Bavarian road infrastructure system) provided by the Bavarian Ministry of the Interior, for Sport and Integration (StMI, 2017) and the Official Topographical-Cartographic Information System Data - Digital Landscape Models (ATKIS-DLM data) by the Surveying Authorities of the Federal States of Germany (Surveying Authorities of the Federal States of Germany, 2015). Weather data from the German weather service (DWD, 2019) were used. The DWD obtains its data from COSMO-REA6 - a regional reanalysis dataset (product of a numerical data model that assimilates terrestrial and remotely sensed weather data) that covers the EUROCOR-DEX area with a spatial resolution of 6 6 km and uses the ERA-INTERIM global reanalysis dataset for boundary conditions. For the land use classification, PANGAEA land use data with a spatial resolution of 30 30 m based on remote sensing Landsat and LUCAS data (Pflugmacher, Rabe, Peters, & Hostert, 2018) from the year 2018 were used in the prediction model.

3.2. Data pre-processing

The different datasets required data preparation and cleaning for the later data mining process. Nearly all data required a trans-

Table 1 Data and derived features as input for the ML predictive analysis (not normalized values).

Data Source Feature/Label Unit Range
Wildlife-vehicle collisions (WVCs)
(spatiotemporal)		 StMI & Bavarian State Police, 2020 WVC label - 1
Non-WVCs
(spatiotemporal)		 Randomly calculated - 0
Road-related ATKIS (mainly) Road class 1–5
(spatial) (Surveying Authorities of the Federal States of Germany (2015), 2015) Road width metre 2 – 22
BAYSIS data (StMI, 2017) Middle barrier - 0–1
Road barriers - 0–1
Additional lanes - 0–1
Ditch - 0–1
Slope - 0–1
Weather Weather data (DWD, 2019) Ambient temperature C 22.20 to 39.07
(spatiotemporal) Precipitation litre/m2 0–72.95
Cloud coverage % 0–100
Land use PANGAEA land use data (Pflugmacher et al., 2018) Artificial land % 0–100
(spatial) Cropland, seasonal % 0–100
Cropland, perennial % 0–100
Forest, broadleaved % 0–100
Forest, coniferous % 0–100
Forest, mixed % 0–100
Shrubland % 0–100
Grassland % 0–100
Bare land % 0–100
Water % 0–100
Wetland % 0–100
Snow and ice % -
Time-related WVC data (StMI & Bavarian State Police, 2020) Day of year - 1–366
(temporal) Weekday - 1–7
Hour of the day - 0–23
Solar altitude 64.50–64.50

formation of geographical projections and data pre-processing to obtain a homogenous key component, with which the features WVCs, and non-WVCs (as label) could be correlated. GIS methods automated by Python scripts were used, namely GDAL and OGR from the OSGeo library (OSGeo, 2019) for interacting with geodata file formats, and shapely (Toblerity, 2020) for geometric processing. The road shapefile ver01_l of the ATKIS-DLM data were divided into point vectors with a distance of 25-metres along the road network to define the set of locations to be examined (Fig. 1). These locations were used as a basis layer to extract the feature values from several other datasets.

From the ATKIS-DLM data, the information about the road class and road width was obtained. Missing values of the road classes were derived from the road names, for example, the federal road class (German: B = Bundesstraße) from the street name ''B 12," while the missing road width were derived from the road classes using standardized values for road construction in Germany (11.5 m for highways, 7.5 m for federal roads, 6.5 m for state roads, and 5.5 m for municipal road (FGSV, 1996). To avoid duplicate information, road axes of highways were deleted as the lane axes are also registered for highways that run parallel to the road axis. The road information was augmented by the street data of BAYSIS. According to this road database, the road network was classified in sections with middle barriers, additional lanes, ditches, and slopes (Table 1). Details in the attribute column legend of the BAYSIS shapefile Straßenbestand.Deckschichtalter were used for the classification (e.g., hard shoulders or bus bays were classified as roads with an additional lane, while protective stripes with crash barriers were classified as sections with barriers). To obtain all infrastructure elements at both sides of the road, vertical lines at the 25 metre distance points from the ATKIS data were constructed. Along these vertical lines, all elements of the BAYSIS shapefile were identified, and the information were written into the 25-metre point layer as key elements for the later data mining.

The land use classes of the PANGAEA data were identified based on the 25-metre points. The land use types were included on both sides of the road within three different radii to test the impact of the environmental land use of an accident within near, medium, and long distances (Fig. 1, semicircles in the right image). The relative area share of each of the 12 different land use classes was calculated separately for both sides of the street within semicircles of 100, 500, and 200 meter radii. Thus, six different buffers were used for each land use type to model the impact of the land use structure at each potential accident location for ML analysis.

Furthermore, the CSV dataset of WVCs was stored as a shapefile and joined to the nearest of the 25-metre points to create a spatial reference. For the temporal reference, the timestamp of the WVCs were discretised to an hourly resolution. In total, 730,867 WVCs were registered but only 567,233 WVCs (78%) were included in the predictive analysis, since the remaining accidents do not contain any geographic references. A set of non-WVC locations was generated in addition to the WVC registrations that are required for the ML approach. A number of 657 billion spatiotemporal points for potential accident occurrences could be obtained by multiplying the 7.5 million 25-meter road points and 87,600 hours of the 10 year study period, but only 12 million non-WVCs are necessary for the latter analysis (see chapter 4). Hence, spatiotemporal points were randomly generated in an hourly resolution based on the 25 meter distance points of the ATKIS road network using the random library in Python based on the Mersenne Twister algorithm, while spatiotemporal locations with the locations and timestamps of the WVCs were excluded to obtain non-WVCs for supervised ML training. For each of the accidents and non-accidents, the timestamp was used to obtain the features weekday, day of the year, and hour of the day. In addition, the solar altitude was calculated for each spatiotemporal point using the solar position algorithm (Holmgren, Hansen, & Mikofski, 2018). Finally, the features air temperature, precipitation, and cloud coverage were extracted from the weather dataset by determining the fitting tile of the weather data at the specific point and at the specific time.

A SQLite database was used to combine the different generated datasets in one dataset with a label (WVC/non-WVC) as input for the ML process. After the spatiotemporal points of non-WVCs and the time-based features were generated, they were stored directly in tables in the database. The same was applied for the weather data, which was selected according to the spatiotemporal locations. The road infrastructure, land use, andWVC data were stored in shapefiles as an intermediate step and, later, imported into the database as separate tables for each shapefile. Finally, each data row has a reference to one of the 25-m points from the ATKIS dataset using WGS 84 UTM system (World Geographical System 84, Universal Transverse Mercator) as geographical reference, and an hourly resolution (the hour of the day and day of the year) as temporal dimension, which are included in all tables as key elements. The normalization of the data was done as part of the SQL statement that joined the several tables to one source table for ML processing. In this table, each spatiotemporal point is characterized by 88 features, whereby-six classes were deleted due to the nonexistence of snow and ice areas (at both roadsides within three radii).

3.3. Case study area

The data were applied for WVC analysis in Bavaria. The Free State of Bavaria in the southeast of Germany covers 7,548 km2

Fig. 1. Data preparation using spatiotemporal points of 25-metres road sections (left image), including, e.g., land use environment within 100-, 500-, and 2000-metres radii (right image).

and seven administrative districts: Lower Bavaria, Upper Bavaria, Lower Franconia, Middle Franconia, Upper Franconia, Upper Palatinate, and Swabia (Fig. 2; Table 2).

Bavaria borders the Czech Republic in the east and Austria in the south-east and south. While a significant share of the around 13 million inhabitants in Bavaria live in a few metropolitan areas (such as Munich or Nuremberg-Fuerth-Erlangen), large parts of Bavaria, such as the Upper Palatinate and Lower Bavaria, are sparsely populated. The northern and eastern parts with the densely forested Bavarian Forest belong to the German low mountain range, while the mostly flat glacial foothills extend from the Alps to the Danube basin. Almost 35% of the Bavarian area is forested. Because of the altitude above sea level and the distance from the sea, Bavaria's climate is rougher and more continental than in most other German federal states. WVCs mainly occur in rural areas, for example, in the county Rottal-Inn in the southeast of Lower Bavaria (Fig. 3). There is a low rate of accidents in the Alpine and forested regions, such as the national parks Bavarian Forest and Berchtesgadener Land, as the road network is not as dense as in other areas. Despite the dense road network, there is a low density of WVCs in urban areas. The WVC density is the highest on state and county roads in relation to the road length (Table 3).

4. Analysis

4.1. Selection of algorithms and parameters

The prediction analysis determines the collision risk in order to warn drivers of WVCs only when necessary. The risk for WVCs was calculated using ML algorithms in Python. The following algorithms were used for the analysis: random forest (Breiman, 2001), feedforward neural networks (McCulloch & Pitts, 1943), and support vector machine classifier (Vapnik, 1963). Since there are no previous studies for risk prediction of WVCs with ML available, several variants of ML algorithms were tested. The algorithms are chosen to test the prediction capability - in contrast to the commonly used statistical hotspot and impact analysis in WVC research (Pagany, 2020). Research on traffic accidents in general already applies the chosen ML algorithms, while there is still a research gap for WVC prediction using such methods. Especially the temporal forecast and the WVC risk prediction in areas without historical datasets about WVCs are an advantage of the ML techniques. In addition, these methods enable a steady training adjustment of the algorithms due to changed environmental conditions. Therewith, dynamic real-time warnings to the driver via a mobile application (wuidi, 2020) is applicable in order to avoid accidents. Hence, the paper focuses on the generation of ML models resulting in high prediction rates through testing.

The impact of different training ratios between accidents and non-accidents on the prediction quality was tested, as the actual risk probability of WVCs and non-WVCs is not entirely known. Only the locations and times with a registered WVC are an indicator for spatiotemporal risk zones, while the remaining road sections and times are assumed to be low risk. While the probability is high that only one, a vehicle, or an animal, or neither of them will cross a certain road at the same time, there is also a possibility that both a vehicle and an animal cross at the same point at the same time, but that the individual reaction of the driver or the animal is such that a collision can be avoided. As the WVC data do not allow any information about these near misses, but the risk is higher than the actual WVC data suggest, higher, and also extreme ratios of WVCs were used for ML training (25%, 50%, 75%, 90% WVCs) (Fig. 4). Afterwards, the models were tested with a uniform ratio of 5% WVCs and 95% non-WVCs to obtain comparable results. Hence, the ML training were conducted with the Bavarian WVC data and different sample sizes of non-WVCs for the predictive risk analysis.

To test the temporal and spatial transferability, WVC data were split in three ways, and the training and test data were enriched with non-WVCs to obtain the abovementioned ratios for training

Fig. 2. Case study area of Bavaria (map on the left side) and districts (map on the right side).

Table 2 Population, area, and road length per districts based on ATKIS data and Destatis, (2020).

| District | Population | | Area | | | Road length | | |------------------|------------|---------|--------|---------|-------------|-------------|--| | | | | km2 | % | km | % | | | Lower Bavaria | 1,192,641 | 9.48% | 10,330 | 14.6% | 44,436,083 | 25.96% | | | Upper Bavaria | 4,418,828 | 35.12% | 17,530 | 24.8% | 27,839,851 | 16.26% | | | Lower Franconia | 1,315,882 | 10.46% | 8,531 | 12.1% | 21,935,317 | 12.82% | | | Middle Franconia | 1,717,670 | 13.65% | 7,245 | 10.3% | 17,027,147 | 9.95% | | | Upper Franconia | 1,067,988 | 8.49% | 7,231 | 10.2% | 18,851,596 | 11.01% | | | Upper Palatinate | 1,081,800 | 8.60% | 9,691 | 13.7% | 17,124,381 | 10.00% | | | Swabia | 1,788,729 | 14.21% | 9,992 | 14.2% | 23,950,389 | 13.99% | | | Total | 12,583,538 | 100.00% | 45,476 | 100.00% | 171,164,764 | 100.00% | |

Fig. 3. Heatmap of WVCs in Bavaria.

Table 3 Distribution of WVCs by road class.

Road class WVC share (%)
Municipal roads 17.09
Country roads 31.43
State roads 33.90
Federal roads 14.88
Highways 2.70

and testing. The random split in a ratio of 80% training data and 20% test data were used as a basis for comparison. Furthermore, the predictive analyses were carried out with a training sample containing WVCs and non-WCVs from between 2010 and 2017, while data from 2018 and 2019 were used for testing (timebased split). Also, the algorithm that was trained on the dataset from Bavaria except Upper Bavaria, was then tested in Upper Bavaria (geographical split) (Fig. 4).

Fig. 4. Analyzing structure for the WVC risk prediction.

From the class of neural networks, the feedforward neural networks performed best in a pre-test, where convolutional, recurrent, and feedforward neural networks were analyzed. Thus, convolutional networks and recurrent networks, which are typically used for image classification and temporal processes, respectively, can be excluded. With support vector machine classifier, a very fundamental algorithm was added. In addition, random forest was chosen as third option to have an algorithm, which yielded good results in previous studies and enables impurity calculation for the input features.

Furthermore, the hyperparameter tuning was done in an automated way using automated ML (AutoML) systems. The StructuredDataClassifier from AutoKeras was used, which is based on the TensorFlow library for the feedforward neural networks (Jin, Song, & Hu, 2019). Moreover, auto-sklearn, which is an AutoML toolkit for the scikit-learn library, was applied for the random forest and the support vector machine classifier (Feurer, Klein, Eggensperger, Springenberg, Blum, & Hutter, 2015). Even though the hyperparameters are chosen by AutoML systems, some specifications needed to be delimited.

In AutoKeras, it is necessary to specify the count of trials and the epochs for the model training. Ten trails and 10 epochs were chosen, as tests with 5 trails were insufficient for adequate results and the training with 15 trails did not improve the results significantly but took an inappropriately long runtime. To determine the number of epochs, several trainings were run with 20 epochs and examined at which number of epochs the accuracy and loss functions did not improve significantly. Using Auto-sklearn, the run time of the entire process and of a single model training was specified, which indirectly defines the number of different hyperparameter settings. In addition, a number of 10 cores were continually used for all model generations. For the further parameters, default values were used.

4.2. Evaluation metrics

To evaluate the results, confusion matrices (Table 4) and further indicators were generated to show the numbers of correct and incorrect matches of the classification. With the accuracy rate, all correctly classified matches are related to the overall number of matches of the confusion matrix (Eq. (1)). The sensitivity, also called recall, is an indicator that shows how often WVCs are correctly classified in relation to the real number of WVCs (Eq. (2)). Specificity shows how often non-WVCs are correctly classified in relation to the real number of non-WVCs (Eq. (3)). While the false negative (FN) rate is the inverse value of the sensitivity rate, the false positive (FP) rate is the inverse of the specificity rate.

Table 4 Confusion Matrix.

| | WVC (1) predicted | Non-WVC (0) predicted | | | |-------|-------------------------------------------------------------------------------|----------------------------------------------------------------------------------------|--|--| | WVC | actual | Actual/true positive (TP),
meaning WVC occurred, and
is predicted to be positive | | | | False | negative (FN), meaning WVC
occurred, and is predicted to be
negative | | | | | Non- | WVC actual | False positive (FP), meaning
no WVC occurred, but is
predicted to be positive | | | | True | negative (TN), meaning no WVC
occurred, and is predicted to be
negative | | | |

accuracy rate = $$\frac{TP + TN}{TP + TN + FP + FN}$$ (1)

$$sensitivity = \frac{TP}{TP + FN} \tag{2}$$

$$specificity = \frac{TN}{TN + FP} \tag{3}$$

Especially high specificity and sensitivity rates are useful indicators for evaluating the quality of a WVC risk prediction model concerning its application as warning service. Therefore, an own measure, the risk prediction quality, is defined as the follows: the mean value of the double-weighted sensitivity and singleweighted specificity rate (Eq. (4)).

$$\textit{risk predictio nquality}(\textit{RPQ}) = \frac{2*\textit{sensitivity} + \textit{specificity}}{3} \tag{4}$$

With the double-weighted sensitivity, the emphasis is laid on the potential to correctly classified WVC risk in contrast to false classified high WVC risk. This sensitivity value is the basis for a warning in the case of WVC risk and only a few non-warnings even if the risk would be high. Too many warnings due to low specificity rates would annoy the drivers without a reason for an accident risk.

5. Results

With accuracy rates between 87.36% and 92.00%, all three ML algorithms assign most observations correctly as WVCs or non-WVCs (bold values in Table 5), using a ratio of 25% WVCs and 75% non-WVCs in the training dataset. With RPQ values of 86.7% and 86.6%, neural networks and the random forest, respectively, performed slightly better than the support vector machine classifier with its highest value of 84.6% (Fig. 5, black line) (Table 6; bold values > 75th quantile). The training ratio of 25% WVCs might achieve the best accuracy rates, but it is not the best choice according to the risk prediction quality measure (see Section 4). While sensitivity rates of under 70% are calculated, more than 30% of the WVCs are predicted as no risk cases, inversely. Using the measure of RPQ, the ratio of 75% WVCs and 25% non-WVCs has been proven as the best training ratio. The training and test run of the different case studies show that it is possible to forecast WVCs in a temporal and in a spatial manner (bold values > 75th quantile; modes: rnd = random, geo = geographical, and time = time-based; Fig. 4). Hence, the applicability of predictive analysis for WVCs is shown.

Based on the random forest, the importance of the individual factors is calculated. The results also show that temporal, environmental and roadside conditions affect the WVC risk (Fig. 6). Espe-

Table 5 Confusion matrix of ML results due to WVC ratio, algorithm (neural network = NN, support vector machine classifier = SVM, random forest = RF), and mode (geographical split = geo; randomly split = rnd; time-based split = time).

WVC non WVC
WVC ratio Algorithm Mode pred.
Non-WVC
= No Warn.
FN		 pred. WVC
= Warning
TP		 pred.
Non-WVC
= No Warn.
TN		 pred. WVC
= Warning
FP		 Accur %
25% NN Geo 45,027 88,600 2,358,856 180,057 91.58
Rnd 34,067 79,379 1,977,663 177,811 90.66
Time 39,280 84,946 1,902,819 161,669 90.82
SVC Geo 47,420 86,207 2,294,631 244,282 89.09
Rnd 39,785 73,661 1,919,229 236,245 87.83
Time 42,538 81,688 1,830,269 234,219 87.36
RF Geo 61,119 72,508 2,386,359 152,554 92.00
Rnd 39,622 73,824 1,953,723 201,751 89.36
Time 41,418 82,808 1,859,100 205,388 88.72
50% NN Geo 19,563 114,064 2,128,044 410,869 83.89
Rnd 14,355 99,091 1,743,486 411,988 81.21
Time 16,158 108,068 1,691,907 372,581 82.24
SVC Geo 21,475 112,152 2,047,878 491,035 80.82
Rnd 16,102 97,344 1,666,100 489,374 77.72
Time 18,361 105,865 1,603,721 460,767 78.11
RF Geo 19,356 114,271 2,087,965 450,948 82.40
Rnd 12,040 101,406 1,638,286 517,188 76.67
Time 14,051 110,175 1,622,192 442,296 79.15
75% NN Geo 7489 126,138 1,812,158 726,755 72.53
Rnd 4867 108,579 1,434,604 720,870 68.01
Time 5503 118,723 1,385,820 678,668 68.74
SVC Geo 7867 125,760 1,666,913 872,000 67.08
Rnd 5766 107,680 1,332,631 822,843 63.48
Time 6116 118,110 1,283,013 781,475 64.02
RF Geo 8546 125,081 1,839,873 699,040 73.52
Rnd 4821 108,625 1,445,009 710,465 68.47
90% Time 5122 119,104 1,388,963 675,525 68.90
NN Geo 1931 131,696 1,305,375 1,233,538 53.77
Rnd 2099 111,347 1,173,934 981,540 56.65
Time 1945 122,281 1,098,623 965,865 55.78
SVC Geo 1506 132,121 1,055,130 1,483,783 44.42
Rnd 1231 112,215 847,635 1,307,839 42.30
Time 1129 123,097 793,250 1,271,238 41.87
RF Geo 2251 131,376 1,348,305 1,190,608 55.37
Rnd 1753 111,693 1,091,939 1,063,535 53.05
Time 1763 122,463 1,053,540 1,010,948 53.73

Fig. 5. Sensitivity and specificity rates and their mean values as indicator for WVC risk prediction.

Table 6 Metrics of WVC risk prediction due to WVC ratio, algorithm (neural network = NN, support vector machine classifier = SVM, random forest = RF), and mode (geographical split = geo; randomly split = rnd; time-based split = time).

WVC ratio Algo-rithm Mode WVC Non-WVC sensitivity FN sensitiv: FN 2
Sensitivity FN
rate
%		 Specificity FP
rate		 þspecificity 2 þFP
2		 2 þ specific: 3
%		 þFP
3
%
% % % % %
25% NN geo 66.30 33.70 92.91 7.09 79.61 20.39 75.17 24.83
rnd 69.97 30.03 91.75 8.25 80.86 19.14 77.23 22.77
time 68.38 31.62 92.17 7.83 80.27 19.73 76.31 23.69
SVC geo 64.51 35.49 90.38 9.62 77.45 22.55 73.13 26.87
rnd 64.93 35.07 89.04 10.96 76.99 23.01 72.97 27.03
time 65.76 34.24 88.65 11.35 77.21 22.79 73.39 26.61
RF geo 54.26 45.74 93.99 6.01 74.13 25.87 67.50 32.50
rnd 65.07 34.93 90.64 9.36 77.86 22.14 73.60 26.40
time 66.66 33.34 90.05 9.95 78.36 21.64 74.46 25.54
50% NN geo 85.36 14.64 83.82 16.18 84.59 15.41 84.85 15.15
rnd 87.35 12.65 80.89 19.11 84.12 15.88 85.19 14.81
time 86.99 13.01 81.95 18.05 84.47 15.53 85.31 14.69
SVC geo 83.93 16.07 80.66 19.34 82.29 17.71 82.84 17.16
rnd 85.81 14.19 77.30 22.70 81.55 18.45 82.97 17.03
time 85.22 14.78 77.68 22.32 81.45 18.55 82.71 17.29
RF geo 85.51 14.49 82.24 17.76 83.88 16.12 84.42 15.58
rnd 89.39 10.61 76.01 23.99 82.70 17.30 84.93 15.07
time 88.69 11.31 78.58 21.42 83.63 16.37 85.32 14.68
75% NN geo 94.40 5.60 71.38 28.62 82.89 17.11 86.72 13.28
rnd 95.71 4.29 66.56 33.44 81.13 18.87 85.99 14.01
time 95.57 4.43 67.13 32.87 81.35 18.65 86.09 13.91
SVC geo 94.11 5.89 65.65 34.35 79.88 20.12 84.63 15.37
rnd 94.92 5.08 61.83 38.17 78.37 21.63 83.89 16.11
time 95.08 4.92 62.15 37.85 78.61 21.39 84.10 15.90
RF geo 93.60 6.40 72.47 27.53 83.04 16.96 86.56 13.44
rnd 95.75 4.25 67.04 32.96 81.39 18.61 86.18 13.82
time 95.88 4.12 67.28 32.72 81.58 18.42 86.34 13.66
90% NN geo 98.55 1.45 51.41 48.59 74.98 25.02 82.84 17.16
rnd 98.15 1.85 54.46 45.54 76.31 23.69 83.59 16.41
time 98.43 1.57 53.22 46.78 75.82 24.18 83.36 16.64
SVC geo 98.87 1.13 41.56 58.44 70.22 29.78 79.77 20.23
rnd 98.91 1.09 39.32 60.68 69.12 30.88 79.05 20.95
time 99.09 0.91 38.42 61.58 68.76 31.24 78.87 21.13
RF geo 98.32 1.68 53.11 46.89 75.71 24.29 83.25 16.75
rnd 98.45 1.55 50.66 49.34 74.56 25.44 82.52 17.48
time 98.58 1.42 51.03 48.97 74.81 25.19 82.73 17.27

cially the road-related factors road width (19%) and road class (14%) are determinants for the WVC prediction. Also, artificial land in the near and middle circle around the potential WVC locations (6% and 4% for each side s1 and s2; see Fig. 6), and time-based factors, such as sun angle (6%), hour (3%), and day of the year (1%), influence the ML prediction. Therein, the importance of the factors should not be understood as individual impact, but as combination of all factors influencing the prediction model. It is also independent of whether the factors increase or reduce the WVC risk. Finally, the ML analysis only determines the collision risk in order to warn drivers of WVCs only when necessary.

6. Discussion

From the large dataset of WVCs, spatiotemporal environmental patterns could be extracted with an ML model that predict the risk for WVCs appropriately. Even when the accident conditions varied case by case, it could be shown that the interaction of different road-. infrastructure-. land use-. weather-. topographic-. and temporal components is a complex cause-effect relationship for actual collisions. With ML and the collected dataset for the last 10 years covering the whole Bavaria, it was possible to predict WVCs spatiotemporally in a resolution and regional extent that had never been achieved before. Especially for dynamic warning systems, highly detailed data mining seems to be a critical advantage for precise warning opportunities. Thus, a detailed space–time-link between the accidents and potential influencing factors was able to be done. To not overlook the broader environment of WVCs, which may have an influence, the proportional shares of each land use type within the near, medium, and long distances, different infrastructure elements, and the topography were considered in the ML analysis.

The case study showed that ML, especially feedforward neural networks and random forests, are suitable for an appropriate risk prediction. Also, the support vector machine classifier achieved a relatively high risk prediction quality (RPQ) – not as high as with the beforementioned algorithms, but within a shorter runtime. As the WVC data does not allow any information about near misses and, hence the actual risk probability is not entirely known, different ratios of WVCs were used for ML training. The models with a training ratio of 75% WVCs and 25% non-WVCs yielded the highest quality of WVC risk classification. AutoML helped to define the best hyperparameters for a high quality of risk prediction.

The results show, on the one hand, the predictability in general, as the RPQ achieved continually values of over 85% using feedforward neural networks and random forests (Table 6). On the other hand, the transferability and applicability of the prediction model is proven, as all modes – randomly split, time-based, and geographical training splits yielded similar high results. The model was successfully tested for the Bavarian region, which shows that a transfer of the risk prediction of WVCs is generally possible.

The developed method can be applied to areas outside the case study region. As WVC data are only available for some areas, having a method that can be applied anywhere may be helpful in tack-

Fig. 6. Impurity results of features used in the random forest model (s1/s2: side 1/2 of the road); impurity 1% is shown.

ling the problem on a global level. However, the global transferability must be proven in future research and probably adapted to regional or species-specific characteristics. Since the case studies were only tested within Bavaria, this hypothesis is limited because it has not been proven whether the spatial transferability between different test areas in Bavaria also applies to other European or global regions. Since the prediction was made primarily with deer (around 80% of the total sample), the system may be overfit by learning pattern recognitions for collisions between deer and vehicles. Hence, the risk prediction was not applied for species that were not as frequently registered within the test region (e.g., foxes or badgers), nor for other non-native animal species. Also, the nonexistence of land use classes such as snow and ice areas in the Bavarian case study area may generate a bias for other areas during the data pre-processing stage. A limiting factor remains the data availability and quality for other parts of Europe and around the world (Pagany, 2020).

Temporal forecast tests showed that it is possible to apply this prediction analysis for dynamic warnings. As the drivers are more precisely warned when the risk is potentially high and animal movement is more likely, warning applications are implicitly also a mitigation measure in the sense of road ecology and wildlife conservation. Real-time warnings regarding the current place, time, and variable conditions avoid familiarization effect for drivers. For providing real time safety warnings, drivers' reaction time as well as vehicle kinematic reaction must be considered, meaning that the warnings needed to be provided earlier for driving along a road. For WVC warning, the application Wildwarner (wuidi, 2020) is already tested successfully with positive feedback from users (Trager, Kalová, Pagany, & Dorner, 2021).

The impurity calculations confirm that static and dynamic factors influence the WVC risk, which is also confirmed in previous studies (Pagany, 2020). Road width and road class especially determine the risk prediction. The influence of both factors was confirmed in several other studies, such as Clevenger, Barrueto, Gunson, Caryl, and Ford (2015), Pagany, Valdes, and Dorner (2020), Santos, Lourenço, Mira, and Beja (2013), or Williams, Collinson, Patterson-Abrolat, Marneweck, and Swanepoel (2019). The majority of the studies identified more open roads with high visibility through wider lanes, more gentle slopes, and fewer curves as reinforcing for an accident risk. Implicitly, higher road classes such as state and federal roads with a higher number of lanes and broader road constructions are riskier than smaller roads. In Germany, WVCs are the lowest on highways due to the high rate of fenced road sections (Pagany, 2020).

In addition, this study also identified artificial or rather urban areas as highly influential for WVC risk. However, the influence of urban areas and population density is not clearly assignable in scientific literature (Pagany, 2020). While some studies analyzed urban areas as riskier for several species (Visintin, van der Ree, & McCarthy, 2017; Wright, Coomber, Bellamy, Perkins, & Mathews, 2020), other studies identified road sections near urban areas as less risky for wildlife (Jakubas, Rys´, & Lazarus, 2018; Kantola, Tracy, Baum, Quinn, & Coulson, 2019; Zuberogoitia et al., 2014). For the WVCs in Bavaria, the WVCs seem to be generally agglomerated in rural areas with high distance to urban areas (Fig. 3).

According to this study, temporal factors - sun angle, hour of the day, and day of the year – also influence the WVC risk. Especially the time factor seems to be related to wildlife's behavior and activities due to migration (Coe et al., 2015), breeding (Grilo, Bissonette, & Santos-Reis, 2009), or rutting (Zuberogoitia et al., 2014). Several studies found daily peaks around dawn and dusk, which is a peak time of movement for animals and for humans (rush hours) (Bíl, Kubecˇek, Sedoník, & Andrášik, 2017; Laliberté & St-Laurent, 2020; Visintin et al., 2016). These peak times also correlates with the changes of light conditions. As examples, Pagany and Dorner (2016) found a temporal peak of accidents with wild boar in the night hours in autumn, while Biggs, Sherwood, Michalak, Hansen, and Bare (2004) identified an increased risk for deer during the early evening hours in autumn and winter. Sullivan (2011) sees the human and animals' visibility, Rodríguez-Morales, Díaz-Varela, and Marey-Pérez (2013) the wildlife's activities at twilight as causal factors. According to their study, Visintin et al. (2016) found an influence of higher temperature on the WVC risk.

The presence or absence of forested areas also influence the risk prediction in the own analysis. In previous studies, the proximity of the road network to forests and cropland is assessed as risky for WVCs. Chen and Wu (2014), for instance, found that threequarters of their investigated WVCs had forest on both sides of the road, while several studies have proven the same increasing effect for cropland (Carvalho-Roel, Iannini-Custódio, & Marçal Júnior, 2019; Cuyckens, Mochi, Vallejos, Perovic, & Biganzoli, 2016; Found & Boyce, 2011).

The importance of the factors should not be understood as individual impact but in the combination of all factors considered into the prediction model. It is not decisive to know the increasing or decreasing impact of an individual factor on the WVC risk, as analyzed by previous WVC research using classical statistics, but to develop a predictive learning model considering the environmental factors altogether.

Targeted traffic and ecological measures are possible through the spatiotemporal data analysis and prediction, for example, to warn drivers dynamically, or to install protective measures only in times of high risk. As the case studies achieved a high quality in the risk prediction of WVCs, a basis is found to apply appropriate measures, on the one hand, for the management of road safety for humans, and, on the other hand, for biological conservation. For humans, the prediction model is able to provide dynamic warnings due to the results with high resolution in space and time. Therefore, impact studies for individual species are not relevant. However, for ecology management, species-specific risk analysis would be interesting, as different species have different activity times. Consequently, measures such as conscious feeding to change the tracks of the animals, should be adapted to reduce road crossings of a species at risky times and zones. As forest and seasonal cropland on both sides of the road led to high WVC risk, measures should be implemented at these particular road sections. Wide state and federal roads should especially be targeted. The effect of light conditions leads us to conclude that temporal warnings could be a suitable method of risk reduction.

Traffic volume and speed were two factors that were not included into the ML analysis, as the available dataset of floating car data does not provide representative information on the actual speed and volume. If the traffic speed and the volume would be available, they might help to improve the prediction performance. Furthermore, it would be interesting to investigate further ML algorithms to improve the WVC risk prediction and, thus, the road safety for human and wildlife. Even when the prediction study tried to include as many conditions as possible, based on extensive data collection and feature engineering, not all eventualities could be considered for the prediction. Even when the dynamic measures, such as the dynamic warning via a mobile application, may mitigate the collision risk as drivers drove more attentively, the remaining risk cannot be ruled out due to negligent driving behavior or unpredictable movements of, for example, startled animals.

7. Conclusions

The paper studied the WVCs of Bavaria from 2010 to 2019 including their environmental conditions, and investigated the performance of a random forest, feedforward neural networks, and a support vector machine for predicting WVC risk in a spatiotemporal manner. A large number of input data concerning WVCs was pre-processed together with a variety of environmental factors and road conditions using GIS. The pre-processing step enabled the researchers to incorporate all components, including the heterogeneous datasets and spatial and temporal dimensions of the chosen criteria into the WVC model. The risk prediction quality - the mean value of the specificity and the doubleweighted sensitivity rate – was calculated as valuable indicator for assessing risk prediction in the event of road accidents with wildlife. With values of 86.7% and 86.6%, neural networks and random forests are the best choice for WVC prediction. A ratio of 75% WVCs and 25% non-WVCs was found to be the best choice for ML training. Support vector machine classifiers achieved acceptable results in shorter runtimes but at lower prediction rates.

The 36 test runs have proven that the approach can be transferred temporally and spatially. The spatial transferability was verified since the algorithm was successfully tested in Upper Bavaria with prediction values as high as with randomly split training data. The WVC forecast in a temporal manner was also proven through training with datasets from the first eight years (2010–2017), and in the prediction and testing with WVC data of 2018 and 2019. Hence, this study shows that a high quality of risk prediction for the occurrence of WVCs is possible, not only for real-time warnings but also for regions where accident data are not available. In addition to previous studies, the tests have also shown that it is important to predict WVCs spatiotemporally, as a combination of static and dynamic factors determines the predictive analysis. Other studies on WVCs have mainly focused on impact and hotspot analysis. This paper extended the WVC research by applying data mining techniques for the WVC risk prediction.

This developed prediction model for WVC risk makes a valuable contribution to take appropriate action at the actual spatiotemporal accident hotspots, even when accidents with wildlife are rarely registered. The model has the potential to be transferable to areas without WVC data availability and, thus, can help decision makers to define better traffic safety control policies, such as dynamical warnings on a global scale. Nevertheless, the limitations of this study show that more efforts should be made to improve data quality and availability of the necessary input data for a globally applicable risk-learning system for the reduction of traffic accidents.

Acknowledgments

This research was funded by the German Federal Ministry of Transport and Digital Infrastructure (BMVI) as part of the mFund project ''WilDa – Dynamic Wildlife-vehicle collision warning, using heterogeneous traffic, accident and environmental data as well as big data concepts" grant number 19F2014A. I thank the Bavarian Ministry of the Interior, for Sport and Integration & Bavarian State Police for the data provision.

References

Abdelwahab, H. T., & Abdel-Aty, M. A. (2001). Development of Artificial Neural Network Models to Predict Driver Injury Severity in Traffic Accidents at Signalized Intersections. Transportation Research Record, 1746. Paper No. 01-

Arevalo, J. E., Honda, W., Arce-Arias, A., & Häger, A. (2017). Spatiotemporal variation of roadkills show mass mortality events for amphibians in a highly trafficked road adjacent to a national park, Costa Rica. Revista de Biología Tropical, 65(4), 1261–1276. https://doi.org/10.15517/rbt.v65i4.27903.

Beshah, T. & Hill, S. (2010). Artificial Intelligence for Development, Papers from the 2010 AAAI Spring Symposium, Technical Report SS-10-01, Stanford, California, USA, March 22-24, 2010. Retrieved March 12, 2020, from http://www.aaai.org/ ocs/index.php/SSS/SSS10/paper/view/1173.

Biggs, J., Sherwood, S., Michalak, S., Hansen, L., & Bare, C. (2004). Animal-related vehicle accidents at the Los Alamos National Laboratory, New Mexico. Southwestern Naturalist, 49(3), 384–394. https://doi.org/10.1894/0038-4909.

Bíl, M., Andrášik, R., Svoboda, T., & Sedoník, J. (2016). The KDE+ software: A tool for effective identification and ranking of animal-vehicle collision hotspots along networks. Landscape Ecology, 31(2), 231–237. https://doi.org/10.1007/s10980- 015-0265-6.

Bíl, M., Kubecˇek, J., Sedoník, J., & Andrášik, R. (2017). Srazenazver.cz: A system for evidence of animal-vehicle collisions along transportation networks. Biological Conservation, 213, 167–174. https://doi.org/10.1016/j.biocon.2017.07.012.

Raphaela Pagany did this research as a senior scientist at the Technische Hochschule Deggendorf. Her research is focused on Geoinformatics, spatio-temporal analysis and risk prediction. In the project WilDa, she developed algorithms for the Wildwarner service. In addition, she did her PhD study at the Paris Lodron University in Salzburg. Currently, she is working in transport planning.