AI and Snow Removal: Clearing the Way for Better Winters

In Montreal, snow is both a pleasure and an inconvenience. While many enjoy watching it fall from the warmth of their homes, dealing with the snowy mess outside, especially when driving, is a different story. Snow removal crews work hard to clear the 10,000 km of streets to keep the city moving and ensure everyone's safety [1]. Snow removal is not as simple as it seems. It's a complex operation that requires careful planning, balancing costs, time, available equipment and, above all, meeting the needs of the public.

As cities grow and demand for mobility increases, snow removal efforts must keep up. One of the biggest challenges is prioritizing which roads and areas to clear first. Traditionally, this is done using rule-based priorities according to the type of roads and the buildings they serve. In reality, the importance of a road can change depending on factors like time of day, traffic, construction work or weather conditions. To make snow removal more efficient, we need to set dynamically adaptive priorities. That means having up-to-date information on the status of the entire road network.

Artificial Intelligence for Snow Removal

Artificial intelligence can revolutionize snow removal by offering real-time insights on road conditions and adjusting removal operations based on on-the-ground necessities. Montreal has an extensive network of traffic cameras, placed at key intersections, capturing images that can provide relevant information for snow removal. By feeding these images into Deep Learning algorithms, we can predict insights such as snow levels and traffic volume to help decide which roads should be cleared first. In this article, we explain how we implement such an approach to enable efficient snow removal operations.

Our method has two main components: one for roads with cameras and another for roads without. In both cases, the goal is to predict the level of snow covering the road, and the volume of traffic.

Roads with and without Cameras

For the first component, we collected and labelled over 40,000 images from Montreal’s traffic cameras during the 2022 and 2023 winter seasons. We then trained a Convolutional Neural Network (CNN) to classify images into four snow cover categories, shown in Figure 1. These categories align with the snow removal operations of the City of Montreal. This prediction helps determine which operations to carry out, and contributes to calculating the priority. The CNN achieves an accuracy of up to 97%, which drops to 79% in extreme weather conditions where visibility is poor. To estimate traffic volume, we count the vehicles passing through each intersection during a set time interval. Using the YOLOv5 object detection model, we identify and count moving vehicles while filtering out parked ones, allowing us to estimate the relative importance of traffic in the road network.

The second component extends the insights gained from camera-equipped roads to the rest of the road network. We represent the road network as a graph, where each node corresponds to an intersection. For streets connected to intersections, we assume they share the same attributes within the camera’s coverage range. However, roads without cameras offer no direct information. To fill in the gaps, we use a graph learning technique called Graph Convolutional Network (GCN). The GCN leverages the spatial structure of the graph to propagate data from camera-monitored nodes to neighbouring intersections and connecting roads, allowing us to extrapolate snow and traffic conditions in areas without camera coverage.

With snow and traffic information now available for all roads, we can set priority levels by combining this dynamic data with other key factors that describe the roads. These include the type of streets (primary, secondary, tertiary or residential) and the type of facilities they serve, such as hospitals, schools, fire stations, etc. These factors ensure that roads providing access to critical services are prioritized accordingly. To simplify snow removal planning, we categorize priorities into four levels, which are used as weights for the road network graph. Figure 2 shows the road priority distribution in the Ville-Marie borough during a snowfall event on February 18, 2022.

Once the priorities are set, the City can apply a routing algorithm to plan snow removal based on these priorities. For the tests, we implemented a routing algorithm that formulates the clearing of high-priority roads first. We then compared it with a standard algorithm that finds the shortest path between two points, regardless of road priority. Table 1 provides an excerpt from simulation results run across different boroughs and snowfall events. Although our method takes more time than the shortest route approach, it prioritizes more critical, high-priority roads. At first glance, opting for the shortest path between two points may seem better because it’s faster and more cost-efficient. However, from a service quality perspective, clearing higher-priority roads is more beneficial, even if it results in slight delays.

Table 1: Operation times (minutes) per routing algorithm.

Although our tests don’t account for all real-world variables, our method offers a promising step toward improving snow removal services, by focusing on where they are most needed at given times. With this approach, we aim to provide municipalities and relevant authorities with a framework ensuring better service for road users, and enhancing quality of life in winter. With the right camera infrastructure in place, this dynamic prioritization method can be integrated into existing routing algorithms, boosting the overall efficiency of snow removal operations.

References:

[1] Ville de Montréal. «Tout savoir sur le déneigement à Montréal». https://montreal.ca/articles/t....

[2] Karaa, Mohamed, Hakim Ghazzai, Yehia Massoud et Lokman Sboui. 2024. « A Computer Vision-Based Framework for Snow Removal Operation Routing ». IEEE Open Journal of Circuits and Systems, vol. 5, p. 81‑91.