Machine Learning for Road Condition Analysis Part 1: Partnerships
Across the world, delivering effective transport infrastructure is key to any nation’s economic growth. It’s necessary to underpin vital services; to transport goods; to enable mobility among a country’s population; and to assist in reducing inequalities across a country. Such issues are particularly prescient to those economies categorised as in need of development aid by the OECD. If you look at that list you might note the number of countries on that list that are in East Africa, a region which is at present facing almost unparalleled mass urbanisation.
Good roads are becoming increasingly important
Many of the cities in countries such as Kenya, Uganda and Tanzania are growing at more than 10% every year. Rural dwellers are flooding into urban areas, in search of better prospects and a better life. This isn’t always what awaits them, and the cities they reach often can’t cope with such a huge influx of citizens, and the unmitigated pressure to public infrastructure that is applied.
In particular, road capacities are being stretched to their limits - if not beyond. In settlements such as Dar es Salaam, Tanzania for example (which was projected to have 10 million inhabitants by 2030 by the World Bank, yet currently looking set to beat that prediction with ease), informal settlements now make up the majority of the city. And as such, roads are not established or planned — they simply emerge. In tandem, better connections to rural areas simultaneously become increasingly important, lest remaining rural communities become even more isolated.
Despite the fundamental importance of having a good road infrastructure, conditions in countries such as Tanzania are often far from optimal both in rural and urban contexts. Countries are not only battling from a relatively low-quality starting point in terms of road quality (especially on low-volume roads) but are also simultaneously facing difficult environmental conditions. In Tanzania in particular floods and heavy bouts of seasonal tropical rainfall are common, continually degrading whatever infrastructure already exists.
Deep Learning to Monitor Road Infrastructure
For the growing economic health of a region such as Tanzania, monitoring road-conditions has therefore become a key issue. And it is a big job. To survey a whole nation, or even just its key road links, is a serious logistical challenge — never mind extending that work to more isolated rural areas. Surveying requires not only significant human resource but skilled application of relatively expensive technologies. Thousands of man hours are necessitated; to slowly traverse tens of thousands of kilometres of roads.
However, a potential solution is in sight — advances in Artificial Intelligence are promising the possibility of automating road condition analysis via “deep learning”. As part of Frontier Technology Livestreaming Programme, a proof of concept study is now underway in Zanzibar to investigate whether this promise can become a reality: “Remote Sensing and Applied Machine Learning for the assessment of Low Volume Road Condition”.
Z-ROADS, as the project is called internally for short, is developing a pilot system, through collaboration between DfID and consultants based at the University of Nottingham, that investigates application of state-of-the-art machine learning techniques to automatically classify road conditions from high resolution drone (UAV) imagery. Our first step is stakeholder discovery: to build partnerships and to ensure that this technology can scale.
Building Partnerships
Projects of this kind, which seek to help accelerate development, must by their very nature be joint affairs. Success can only come from shared ownership of the project and through the formation of real partnerships — collaborations between developers, researchers and the domain experts. In the case of Z-ROADS, Sprint 1 has seen partnerships rapidly established, thanks to fantastic engagement by Zanzibar’s Department of Roads (DoR) and Zanzibar Road Fund Board (ZFRB).
For Z-ROADS, two initial meetings with DoR were followed by a successful project inception workshop - attended by the project team, as well as representatives from the Zanzibar Road Fund Board, DROMAS (a District Road Management System used within Tanzania) and the Dar es Salaam Institute of Technology. This, combined with presentations to the Tanzania Road Fund Board, have established some great engagement, and has seen the rapid approval for the project’s data collection and quality control protocols by ZFRB.
In fact, through these partnerships the project was rapidly able to establish a final extent for its survey regime (see the figure below for a graphical illustration of planned road coverage, ready to be circumnavigated in sensor laden vehicles). This is essential stuff for obtaining the ‘ground truths’ — real world road condition scores, that we can label roads with, and which we can use to train and test our Artificial Ingelligence models with. It was agreed that the project would attempt to cover the whole of the DoR road inventory to obtain a full set of scores for our test region of Zanzibar (~700km of unique roadways, including both paved and unpaved roads).
Local Knowledge is Invaluable
Further still, from partnerships comes invaluable advice. It is all to easy for projects, especially those established and funded in western arenas, to swing in with a “we know best” attitude combined with a bucket full of technological solutions. We have learnt (sometimes the hard way) that on development projects it is people on the ground who know the most about their own contexts - and, in particular, the idiosyncrasies a project will face given the geographies in which it is situated. Partners aren’t there to just give credence and support to the work, but to help guide the plan. And in this case, we learnt that while machine learning is a fine thing, it is nought in the face of the elements.
The advice of our local partners was “survey now and survey quick”. Rainy season in Tanzania is a serious affair — torrential rain deluges the region, and regularly can continue unabated for a week. Every year this brings with it natural floods, which wreak havoc on roadways in particular. In urban areas especially, which often provide insufficient drainage due to inadequate urban planning, these floods become man-made disasters. Surveying in such conditions is simply not possible.
In light of this, the program was accelerated, with surveying being rapidly brought forward in order for it to be completed by December, 2017 - thanks to the flexibility of FTL’s program setup and through the collaboration of DfID and DoR with the project team. Due to the partnerships formed it was possible to rapidly bring the required equipment and skills together, and the team were able to begin the surveying process before conditions became unfavourable. In a DoR vehicle, fitted with both sensors and team members, data collection began in earnest.
This process actually achieved excellent results within two weeks… but data challenges from other data sources were ahead — more of which in the next part of this blog:
