Part 2 - What can AI see?
The Problem
Globally, 2.2 billion people lack access to safe drinking water, and over half of the world’s population does not have access to safe sanitation (UNICEF, n.d.). Nepal has a vast Network of WASH infrastructure (pumps, containers, pipes, etc,) to provide access to clean water. The task of maintaining this infrastructure is immense. Over time, cracks in pipes lead to leaks, water pumps become deficient, and all these issues compromise both the cleanliness and the supply of water to communities across the country.
To facilitate better management of WASH assets, the Ministry of Water Supply, recruited a decentralised network of enumerators to catalogue the condition of WASH infrastructure in their communities. This data was collated into the NWASH portal. A database which included a vast number of images, labelled with different faults in the WASH assets. The Ministry intended to use this data as a basis to support local governments to create WASH plans and prioritise resources to the areas where people needed them most.
However, the enumerators gathering the data were not WASH experts, and as such some of the images collected were mislabelled and inaccurate. A team of WASH experts were recruited to identify the correct faults in the collected data. With over 1 million images in the NWASH portal, the amount of data that needed to be processed, far exceeded what the team could review.
The Idea
The Department for Water Supply and Sewerage Management, in partnership with FCDO and the Frontier Tech Hub, recognised an opportunity to address this gap by implementing a computer vision system. The aim was to develop a system capable of analysing the database and detecting infrastructural faults in images not reflected in the textual data, particularly in areas where discrepancies exist between textual data and photographic evidence.
By achieving a more accurate database, experts could streamline the data validation process, thereby enhancing planning efficiency and mitigating risks associated with inaccurate information, ultimately leading to better planning outcomes.
Assessing whether computer vision is a good fit for a particular problem can be challenging.
As we will explore later in the module, there are a few criteria that give a good indication that computer vision is worth exploring. The problem Rara Labs were exploring highlights three key criteria:
-
While there were inaccuracies in the labelling of the images, the NWASH portal collated a huge amount of context-relevant images which could be leveraged to train the model.
-
The existing process already involved a high level of error, as such there was scope for AI to increase the level of accuracy and improve the effectiveness of the workflow
-
AI solutions are particularly valuable in contexts where you have low access to human resources and expertise. Automation allows you to analyse huge datasets that would require a significant amount of human labour to process. Additionally, by training a computer vision system on expert input, you can democratise access to that expertise through the application. This leaves the people you do have working on the solution more time to attend to higher-level strategic tasks, in this context, creating effective WASH plans to support local governments.