More efficient monitoring of unpaved road conditions using satellite imagery and artificial intelligence
InfraTech enabling approach(es)
TRL has created an efficient solution for assessing the condition of unpaved roads via high-resolution optical satellite imagery that are processed using artificial intelligence (AI) to categorise the condition of each length of unpaved road on a network, enabling roads to be prioritised for further inspection or treatment. The machine learning assessments performed by the AI have been shown to be more rapid and consistent than manual assessments.
TRL’s solution was developed in six countries in Africa by working closely with road authorities and leveraging experience of developing the UK’s industry survey standards. The solution was one of the top 5 selected by the Global Infrastructure Hub and its eight MDB partners in the Call for Submissions for Sustainable Roads in Emerging Markets.
Need and use case for the technology
Poorly maintained rural roads limit economic development, perpetuate poverty, and limit the physical and social health of communities. This problem is most acute in Africa where over half of the population relies on unpaved roads.
Local road authorities (LRAs) in low- and middle-income countries (LMICs) have limited resources to gather information on unpaved rural roads, and traditional methods of collecting this information are slow and unreliable. As a result, most have limited knowledge of the condition of their unpaved rural roads. This negatively impacts on the planning and prioritisation of road maintenance.
TRL’s technology provides LRAs with a cost effective, practical method to address this problem. The TRL solution utilises high-resolution optical satellite imagery, which are processed using AI to rank the condition of lengths of road across entire road networks.
By providing LRAs with access to previously unavailable highway inventory and condition data on their networks, the solution enables a step-change from the status quo of low-data, low technology, and poor quality management of unpaved roads to a data-driven system that is optimised for the developing world to make better use of financial and physical resources.
Deployment of the technology in Africa – where the African Development Bank (2014) estimates that 54% of roads are unpaved – could be transformative. Widescale implementation would bring significant benefits to the management of a large proportion of the current highway asset. The solution is also compatible with GIS based asset management systems, allowing LRAs to take full advantage of the data collected.
Overview of the technology
The optical satellite images used by TRL are acquired from standard orbiting satellites, which can be tasked with providing specific imagery. The AI solutions used to process the images include customised and blended machine learning techniques. The condition data generated are used by local road engineers to prioritise the roads for further detailed inspection or treatment.
There are a wide range of unpaved roads in developing countries, each managed to meet local priorities and maintained to local budgets and resources. Recognising this, TRL has enabled its solution to be customised to local need and capacity, using the priorities and standards of the LRA are foundational inputs to ensure a sustainable solution.
The machine learning model is calibrated to ensure assessments are in accordance with local materials, construction methods, and maintenance standards. Parameters for defect types and severity ratings are established through the collection of ground truth information on a representative sample of roads, using low-cost equipment that can be deployed in an LRA’s vehicle to collect visual condition and roughness information. Once the model is calibrated, it is applied to the satellite imagery to produce a high-level assessment that applies the LRA’s rating scale for roads.
TRL’s solution can determine the condition of unpaved networks at a high level rapidly, objectively, and with minimal presence on the road – eliminating the cost, logistics, and delays of extensive driven surveys.
Overview of the company
TRL is a consultancy that specialises in research, innovation, and advice. It serves a range of clients in the UK and internationally and has a pedigree of over 85 years as the UK’s centre of excellence in transport infrastructure and services, and related research activities. TRL is a non-profit distributing company, which means that any profits made are reinvested into research and building the company’s capacity as a specialist research organisation.
TRL has helped to shape the modern road environment by developing standards, manuals, guides, and research notes in all aspects of highways, bridges, and transportation. Since 1955, TRL has also supported international assistance in the highways sector by helping the Global South to establish and maintain transport infrastructure. This work now extends to all forms of transportation, environmental, and safety advice, and more recently climate mitigation and resilience. In support of this work, TRL has developed a wide range of software and hardware products, including its own Road Asset Management System (RAMS) that is managed by its sister company, TRL Software. TRL has completed more than 100 major projects in LMICs over the past 65 years.
Transformative outcomes of the technology
TRL’s solution has several transformative outcomes for the roads sector:
Reduced emissions. TRL’s solution minimises physical visits to the roads (visits are needed for calibration only), saving emissions compared with traditional driven surveys. Also, better road conditions lead to more efficient use and shorter travel times – and therefore lower emissions.
Accessibility and affordability. The solution is designed to produce more efficient transport services and greater accessibility. A consequence of this is lower travel costs for local users and less time spent travelling.
Economic growth. There is plenty of evidence that improved rural access can enable economic growth,with investment in rural roads providing better returns than other forms of public expenditure. More efficient rural roads lead to positive impacts associated with increased income, employment, agricultural output, education, and health. Timely maintenance of the roads reduces vehicle operating costs and tariffs for travel to and from rural areas. Most people in rural areas rely on public or informal transport, which reacts to key factors such as demand, accessibility, and operating costs, so improved roads will benefit these users. Agriculture is also likely to benefit as produce will be able to get to market more quickly and with less damage through smoother roads. Agricultural producers can see significant proportions of their highly perishable goods wasted when roads are rough and poorly maintained.
Improved safety. Although road crashes tend to be less serious on poor roads with lower speeds, there are still risks for travellers. This includes vehicles leaving the carriageway when the road becomes slippery or damaged, especially in hilly or mountainous regions. Motorcycle use has increased exponentially in recent years and motorcyclists are at particular risk on roads with poor surfaces, especially when they are overloaded. The data collected by TRL’s solution can help with assessing road safety risk, but the imagery can also be used to identify geometric issues with the road and help plan new alignments where necessary.
Disaster and climate adaptation. These are a challenge for unpaved roads, as they are more vulnerable and generally receive lower levels of maintenance. TRL’s solution can help to identify the most vulnerable areas of roads. The imagery used in this solution can also be used to identify and monitor off-road hazards such as landslides, erosion, and potentially damaging farming practices.
More accurate monitoring of access. The Rural Access Index (RAI), now incorporated as SDG9.1.1, measures the proportion of the rural population with access to an all-season road. TRL refined the methodology for SDG 9.1.1 to make it more sustainable and accurate, and the most difficult part of the RAI is to measure all-season access. This was traditionally done by using surface type, or a combination of condition and surface, which was often not readily available. The use of satellite imagery, along with open source flooding maps and other remote sensing data, could help to measure SDG9.1.1 more accurately and sustainably to ultimately improve access.
TRL’s solution has been developed through application in six countries in Africa:
Nigeria. A research project funded by Satellite Applications Catapult explored to what extent satellite imagery could be used to increase knowledge of the rural road network in Nigeria. This identified the primary indicators of condition and the information that could be gained from satellite imagery. Overall accuracy to match road conditions to the actual condition (as measured on a separate project) was 63%.
Ghana, Kenya, Uganda, and Zambia. This UK Foreign, Commonwealth and Development Office project extended the Nigeria research and further defined the knowledge and benefits that could be gained from TRL’s solution, including identifying unpaved roads as the focus of the process and summarising the potential costs and benefits. The deterioration on earth and gravel roads was clearly visible through differences in surface texture and variation in road width. This process was undertaken manually using experienced engineers, and accuracies of between 63% and 85% were achieved, depending on the categories of condition employed.
Tanzania. This extension of the previous project was undertaken for proof of concept, to demonstrate and further refine the process, as well as considering the institutional and practical implications of using the technology. TRL worked closely with local engineers to identify and ground truth roads, and to understand the practical and institutional issues that could provide barriers to the process.
Tanzania. TRL funded another project to apply machine learning to the process to test the potential for automation, with the aim of making it more objective, faster, and more reliable. Machine learning was focused on the pixel variation on the road surface, with the principle that a higher variation would indicate a poorer condition road. The machine learning results exceeded manual assessments of pixel variation in terms of accuracy, using a confusion matrix approach, with results of up to 89% accuracy to predict conditions for a blended model using classical machine learning.
Tanzania. TRL trialled lower-resolution imagery to reduce the capital costs of the technology. Using 0.75m to 1.0m spatial resolution imagery instead of 0.3m to 0.5 m resolution can reduce the capital costs of the imagery by up to 75%. A trial was designed and undertaken to apply machine learning assessments to the images using the lower resolution imagery (produced by downsampling the original imagery of Tanzania to 1.0m spatial resolution). The results for a blended model from a representative training data set produced an accuracy of up to 85%, but when trialled on ‘unseen’ roads (roads that were not used to train the machine learning) the accuracy reduced, suggesting that the model was overfitting. A project to extend this trial and improve the results for lower resolution imagery has just started, funded by the European Space Agency.
Next developments of this technology
To continue the development and implementation of its solution, TRL is currently exploring:
Options to enhance the practicality and affordability of image sourcing, including through local partnerships
Additional trials to refine the machine learning models and continue demonstrating the value of the solution
Ways to better understand the needs and barriers a country has to asset management, to continue refining the solution.