Adequate water resources along the road provide a suitable environment and resources for seed germination and early seedling growth for invasive species. After the early establishment and the naturalized species overcoming different stresses to produce seeds, the rapid spread of the reproductive offspring makes the species invasive . Human-assisted dispersal potentially creates a longer-range spread than the dispersal mechanisms related to species’ reproductive traits. According to Mortensen et al. , human activities are the main facilitators of the weedy and invasive species spread, and their study indicated that paved roads could spread more weeds than forests and wetlands. Specifically, vehicles are the spreading vectors of long-distance dispersal for weedy and invasive species. The traditional species dispersal model described LDD as a rare event; most cases are seed dispersal by animals , where weed seeds are adhesive to animal fur and travel along with seasonal migration . However, Nathan proposed that human-mediated LDD has become the most important mechanism of LDD in plants and animals, which is a challenge for future LDD prediction. According to Baker , no specific weedy traits or natural dispersal mechanism to help invasive species overcome the large-scale geographical barrier. Nevertheless, human-assisted dispersal can potentially transfer seeds over 100 km away from the parent plants. The 100 km was an approximated cut-off to classify plants as alien species in the model proposed by Richardson et al. . A study in Germany collected seeds from roadside verges, cannabis dryer and their results indicated that nearly 30% of the species collected were by LDD and some species they identified are highly invasive in other countries .

Seeds dispersed by vehicles share common characteristics that might facilitate car-borne dispersal. Zwaenepoel et al. provided another perspective by collecting seed samples from mud attached to the car; the results suggested that carborne floras were pioneer species with small and light seeds. Other traits like large seed production and the ability to reproduce vegetatively are also reported in many studies . Also, a systematic review summarized that about 626 species in 75 families had been identified from cars, and Poaceae is the most frequent family , followed by Asteraceae and Fabaceae . The spread of invasive and weedy species along the road in the real world could be more significant than what has been reported in scientific studies. The large-scale spread of Microstegium vimineum was reported by observation; however, according to a model prediction, the species spread only by natural dispersal is limited . The contrasting results support that human-assisted dispersal leads to an increasing spread rate, and more resources and efforts should be put to roadside vegetation management. Roadside vegetation management is on a large scale, and the local government is the primary agent for management. For example, in California, the California Department of Transportation manages the vegetation along California state highways, and develops projects to protect motorists, cyclists, and potential wildfire spread along the roads . Compared to agricultural weed management, roadside vegetation management has limited tools. The most common practice is mowing, but it requires multiple applications in a short period; therefore, mowing is expensive and ineffective since only the foliar part of the plant is damaged . Herbicide application is used with mowing as the Integrated Vegetation Management . Chemical control could be effective under roadside conditions. For example, herbicide trials conducted in six different regions of Indiana demonstrated that herbicide application could effectively control broad leaf species for more than one year and grasses for months .

Herbicides can significantly lower the cost, but the increasing herbicide-resistant population is another potential concern for roadside weed management . IVM program is important to manage roadside invasive species, and similar to Integrated Weed Management, early detection and monitoring are also main components in IVM. The first step in studying and analyzing the ecological aspect of a specific invasive or weedy species is to conduct a species survey and map the population distribution.A species distribution map is a common approach for evaluating the extent of plant invasion and provides a baseline for the informed allocation of resources and efforts. Botanists usually conduct field surveys to collect plant species, including weedy species. The benefit of this detailed survey is the high accuracy of the species identification and location data, but a detailed survey requires enormous resources. For example, a county-level survey of 3000 km required 35 months, and the researcher had to travel by car, on foot, or even by boat . A typical 3000 km survey is considered small-scale but still time-consuming, labor-intensive, and requires equipment like a vehicle and an accurate GPS positioning system. A field survey is reasonable for species local population examination and species-environment interaction analysis. A car survey can be rapid by applying different sampling or examination methods. According to Shuster et al. , a car survey has a similar probability of finding Alliaria petiolata compared to a survey on foot but requiring four times fewer person-hours. The car survey can involve transects or random sampling sites along the roads base on land uses, soil types, rainfall, and vegetation types . The data collected from the car survey can be used to build a model to understand the relationship between species distribution and environmental factors. A car survey could yield consistent results when various factors are examined in the experimental design. For instance, the traveling speed can vary for different types of roads. For highways, the driver must drive above the minimum speed so that a higher speed can result in lower identification accuracy. Observation accuracy is another factor that affects data consistency. Catry et al. conducted accuracy tests to evaluate the potential human errors. However, it is challenging to include human error rates in the species distribution maps, and in most cases, the human errors in the car survey are unaccountable . A standard methodology for roadside species surveys should be established to yield consistent and comparable data. Some studies do not include all road types in the car survey to save time and reduce costs. For example, Catry et al. excluded the highway or freeway because they believed that the roadsides of the freeway are well-managed and there will be less possibility of having invasive species populations. A survey for all roads in a state will take an unrealistic time to complete, and the cost of traveling will be expensive. For example, California has about 622,000 km of roads, and it will take about 780 days if a driver drives 800 km per day . Thus, government agencies will hesitate to conduct surveys because of limited funding and resources. However, roadside vegetation assessment can help identify the level of invasion and the potential damage. A car survey is not cost-effective for accomplishing a quick assessment on a large scale. Furthermore, a species map can be used in large-scale species dispersal models in which the input data are usually from global or regional databases. Kadmon et al. argued that since randomized surveys on a large scale are rare, those models often rely on incomplete databases with biased data, such as herbaria, natural museums, and user-uploaded entries. The unified database includes data collected from different observers, and we cannot estimate the potential human error if we rely on these databases to run the ecological models. As a result, we need a more systematic approach for large-scale species surveys.Google Street View is a tool in Google Maps and Google Earth that allows users to interact with the panoramas along streets and roads in many countries. GSV and Google Earth are well-developed and well-maintained databases, and the imagery has been online for more than ten years.

Google sends numerous data-collection vehicles on the roads, and those vehicles take 360-degree photos, e.g., cannabis growing systems by installing a rosette camera on the top of the car . According to Anguelov et al. , this project aims to organize a large amount of information, and billions of users can have access to that information. GSV is well-known among ordinary users for educational and recreational activities, but now, these images can be used for ecological studies and vegetation management. For example, GSV was used to map the distribution of the Pine Processionary Moth , in an area of about 45,000 km2 . Their research suggested that this method can be effective if the target species are distinguishable in the GSV images. They also mentioned that the coverage area of GSV still needed to be completed back in 2013. In the past ten years, GSV has been used in several ecological studies. For example, according to Hardion et al. , the integration of ground and aerial images could create a better species distribution map of giant cane , a common grass species along the road, and the species distribution data produced better results in species distribution models compared to the traditional field survey. Additionally, two studies compared GSV and field surveys by car. Studies by Deus et al. and Kotowska et al. reported that the results produced by GSV resemble that produced by the field survey. The two studies surveyed different species of different sizes . Although GSV is more cost-effective than car surveys, studies discussed above have used human observers for plant detection, which is tedious, time-consuming, and not feasible for large-scale mapping. Furthermore, most previous studies reported some common limitations of this imagery database. Most studies suggested seasonality and time differences among the GSV images. The images from different areas were taken in different years or seasons and at different times during the day , which could impact the presence and absence of the survey species or detectability of species within the GSV images . Deus et al. also reported that contrast, ambient light, and sharpness would affect the identification accuracy in some images. Another issue with GSV is that most species were undetectable in their seedling stage when they were small and did not develop distinguishable features . GSV database has been successfully used as a cost-effective method in terms of time and resources, and this benefit can allow researchers to conduct a comparatively large-scale survey in a short period .Artificial Intelligence is a powerful tool to automate tasks, providing insightful solutions for scientific studies in different fields. For example, image detection built with deep learning algorithms can be used in plant species identification. Computer vision methods are successfully used in crop and weed classification to build robotic machines to conduct real-time weed detection in the field . Machine learning includes two main types: unsupervised and supervised learning. Unsupervised learning trains models with unlabeled data, while supervised learning requires labeled data, and most neural networks are examples of supervised learning algorithms . Artificial neural networks are models built based on the nervous system of the living vertebrate . In recent years, neural networks have evolved into the convolutional neural network , which consists of multiple convolutional layers. Many studies have proven that CNNs can reach high accuracy in object detection in images . Dyrmann et al. trained a CNN model to classify young seedlings of 22 species with the average accuracy of 86.2% under controlled indoor conditions. For example, Sugar beet , barley , and Thale Cress could achieve 97% to 98% accuracy . Dyrmann et al. tested another CNN model called DetectNet in overhead images from highly occluded cereal fields, and the model had a recall of 46.3% and a precision of 86.6% for detecting the non-crop plants. The high-occluded fields resemble roadside environments where plants overlap each other. Another well-known fast-detection CNN model called You Only Look Once can be used to detect weeds in outdoor and natural light conditions. Dang et al. tested different versions of YOLO detectors on 12 weed classes at different growth stages in cotton fields. The precisions of 12 weed classes ranged from 81.5% to 98.28%, and the recalls ranged from 78.62% to 97.9% on the YOLOv3 detector . The examples above are all based on overhead images in crop fields, but only a few similar studies have been done under roadside conditions. One example of the roadside condition is a study that integrated GSV Imagery and a CNN network to map the distribution of different crops along the roads and achieved accuracy levels of 92% in California and 98% in Illinois .