Predicted road speeds cut route times by 12% in Sumatra

Summary

• In the Sumatra case study, fewer than 5% of roads had speed-limit information, making it difficult to estimate realistic journey times from existing map data alone.

  
  

• Using satellite images, GPS traces, rainfall and time of day, the model estimated how fast vehicles could travel even on roads with no direct GPS observations. Its predictions differed from observed speeds by about 8.5 km/h on average.

  
  

• When those estimates were used to choose the fastest route rather than simply the shortest one, journeys were 12% faster on average. Nearly one in four routes reduced travel time by more than 20%.

Why is road-speed information difficult to use in routing?

Digital routing depends on more than knowing where a road is. It also needs a realistic estimate of how long that road takes to travel. In many rural areas, speed limits, traffic patterns and consistent GPS observations are limited.

Conditions also matter. Rain can slow travel on unpaved roads, while darkness can change achievable speeds. A route that is shortest by distance may therefore not be fastest in practice.

Why is this a prescriptive analytics problem?

Route planning is a decision problem: given several possible roads, which route should a vehicle take? That decision depends on estimates of travel time.

This project linked prediction with prescription. It first estimates road speeds where direct observations are sparse, then uses those estimates to choose routes based on expected travel time rather than distance alone.

How did the team approach the problem?

The study was developed by Valentijn Stienen and Joris Wagenaar at Tilburg University, Dick den Hertog at the University of Amsterdam, and Jan F. de Zegher at MIT, in collaboration with PemPem, which operates a digital marketplace used by smallholder farmers in Indonesia.

The model combines GPS trajectories, open satellite imagery, recent rainfall and time of day. Satellite images provide information related to road width, surface and surrounding terrain, while GPS observations show how vehicles actually moved through parts of the network.

The model learns patterns from roads with observations and applies those patterns to roads where no tracked vehicle appears in the data.

What did the findings show?

In Sumatra, the model predicted travel speeds with an average error of about 8.5 km/h. The team then compared routes selected by shortest distance with routes selected using predicted travel time.

Travel-time-based routing reduced journey time by 12% on average. For nearly one in four routes, the reduction exceeded 20%.

While the team did not measure income, health or service-delivery outcomes directly, the operational result was that organisations using the same road network can make routing decisions with a more realistic estimate of how conditions affect travel. For smallholder farmers, logistics operators and health services, shorter travel times can mean lower transport time, more predictable market access and fewer delays in moving time-sensitive goods or supplies.

How can the findings inform other applications?

The method uses open satellite imagery and GPS data that organisations may already collect, so it can be tested in other settings where road-speed information is incomplete.

The same estimates could inform delivery planning, emergency logistics and decisions about where services such as health facilities should be located. Future applications could test how well the method transfers across different road surfaces, climates and transport systems.