Data for Children Collaborative

View Original

Building Heights Phase 2 - Calculating Line-Of-Sight Between a School Location and Radio Antennae Using Deep Learning Models on Satellite Data

Calculating Line-Of-Sight Between a School Location and Radio Antennae Using Deep Learning Models on Satellite Data

The Issue

Half of the world’s population has no regular access to the Internet. Millions of children leave school without any digital skills, making it much more difficult for them to thrive and contribute to local and global economies. This has created a digital divide between those who are connected and those who are not, a divide that has become even wider during the COVID-19 pandemic. 

Why Does it Matter?

Article 28 of the Convention on the Rights of the Child states that every child has the right to an education. Similarly, the UN Sustainable Development Goals have a specific focus to ensure inclusive and equitable quality education and promote lifelong learning opportunities for all. However, it is estimated that 2.7 billion people are still offline with 96% of these people living in developing countries. This lack of connectivity means that children have fewer opportunities to learn and fulfil their potential1.   

Providing internet connectivity within schools is essential to enabling the development of digital skills and providing access to online learning opportunities. In addition, connecting schools enables connectivity of local communities and the provision of wider services1.  

In 2019, UNICEF and the International Telecommunications Union (ITU) launched Giga, a joint initiative with an ambitious goal of connecting every school in the world to the internet. A major component of this programme is bridging the connectivity investment gap, including the provision of data-driven insights that support sustainable internet infrastructure investment.  

Our Project

To enable investment in sustainable internet connectivity, the Giga programme works closely with governments to present commercially viable projects. When building their business case, it is essential to understand the existing infrastructure, including line-of-sight between schools and the nearest radio antenna. This information is, in general, not readily available. Put simply, Giga need to know if there are buildings or terrain higher than the school in a straight line between the school site and the chosen antenna.  

 

There are well established tools to calculate line-of-sight between two locations which take Digital Surface Models (DSM) into account. These models typically account for naturally occurring objects and geographic accidents. However, as there is very little available data on building heights, a DSM in isolation is not adequate to calculate actual line-of-sight. 

 

For phase 1 of this project, the Data for Children Collaborative partnered with the Pivigo S2DS programme. The programme uses data challenges to turn practicing academics into data scientists. Working over an intensive five-week sprint, the team developed a deep learning model that could calculate building height from earth observation satellite data, creating a Digital Building Model. The team were able to demonstrate that it is possible to train a convolutional/deconvolutional neural network to learn building heights from high-resolution Maxar satellite data. 

In phase 2, three of the existing S2DS team worked with Data for Children Collaborative and UNICEF to address challenges identified during phase 1 while also developing techniques that adopted medium-resolution Sentinel 1 and Sentinel 2 satellite data. Enabling the use of medium-resolution (and openly available) data ensures the model is cost-effective to develop and scalable in the long term.  

The team focused on building and testing the model on the São Paulo region in Brazil, used during phase 1: 

  1. Addressing a vegetation bias, enabling the model to detect and exclude areas of high vegetation from training datasets. 

  2. Exploring model performance when including Sentinal-2 and Sentinal-1 satellite data and removing costly high-resolution MAXAR data. 

  3. Investigating the capability of non-deep learning techniques for predicting building heights (i.e. Linear Regression, Random Forest, and XGBoost)  

 

As an outcome, the team were able to establish the optimum combination of Sentinal-1 and Sentinal-2 data components to enrich the deep-learning model and produce a more sustainable, cost-effective output.   

 

Theme

Education

Who is involved

 


Our Outputs

See this gallery in the original post