A team of researchers has unveiled a pioneering method for accurately detecting vehicles and estimating their speeds using medium-resolution satellite imagery, promising a scalable, cost-effective solution for traffic management, both in dense urban environments and across large geographies.

The researchers, from Germany’s Heidelberg Institute for Geoinformation Technology (HeiGIT), and Poland’s Institute of Urban Geography, University of Lodz, have tested the method over highways across Germany and Poland, with promising results compared to established methods of traffic monitoring such as road cameras and monitors.

Current approaches to monitoring traffic flow often face significant hurdles related to coverage, cost, and privacy. Traditional stationary methods, such as loop detectors and roadside cameras, offer precise, localised data but are fixed to specific locations and suffer from high installation and operational expenses, making city-wide deployment challenging.

Mobile systems, like Unmanned Aerial Vehicles (UAVs), are limited in the area they can cover and are often constrained by legal restrictions. GPS-equipped probe vehicles, while offering broad coverage, raise major privacy concerns because they rely on collecting and analysing user location data. The penetration rate of GPS devices also remains low, resulting in data sparsity and potential biases.

Previous attempts to use satellite technology, particularly high-resolution satellite video, have proved costly, lacked scalability, and were prone to errors due to low frame rates.

The new research overcomes these limitations by employing medium-resolution PlanetScope SuperDove imagery (approximately 3.7 m GSD), which provides global and daily coverage. According to the researchers, this approach is inherently scalable and independent of ground infrastructure.

The major advantage of the proposed method lies in its focus on a vehicle’s spatial structure rather than relying solely on spectral contrast. By analysing the subtle band timing differences inherent in the satellite’s push-broom sensor, moving vehicles leave behind an effect known as the “object in motion echo”: a blurry, elongated cluster of pixels reflecting the vehicle’s position across the sequential blue, red, and green band captures.

To detect and track these subtle movements, the researchers employed a Keypoint R-CNN model, a type of Convolutional Neural Network (CNN). This deep learning model is configured as a pose estimation technique, trained to identify three connected keypoints that represent the vehicle’s geometric trajectory across the RGB bands.

The ability of the Keypoint R-CNN to accurately model this geometric consistency and movement allows the use of lower resolution satellite images (like the 3.7 m PlanetScope data) effectively. This is significant because scaling methods based on very high-resolution images remains challenging. By focusing on tracking key geometric points, the CNN-based method ensures consistent motion detection even in complex environments or when the vehicle’s spectral difference from the road surface is minimal.

This scalable technology holds significant potential for cities and regions lacking comprehensive data. According to the Heidelberg researchers, it can play a vital role in optimising traffic flow, mitigating congestion, and enhancing road safety. The daily, near-global coverage offered by SuperDove imagery is particularly valuable for macro-level network monitoring, identifying under-serviced areas, and filling spatial and temporal data gaps left by traditional sensors and GPS probes.

While the method demonstrated feasibility in detecting and tracking vehicles at moderate speeds, achieving a Mean Average Precision (mAP) of approximately 0.53, challenges remain. Validation showed a systematic underestimation of speeds for fast-moving vehicles (above 100 km/h). Despite these limitations, the research has established a significant new path for cost-effective, broad-range traffic monitoring.

The paper, “Deep learning enhanced road traffic analysis: Scalable vehicle detection and velocity estimation using PlanetScope imagery” has been published in The International Journal of Applied Earth Observation and Geoinformation, and is available online: https://doi.org/10.1016/j.jag.2025.104707