Will we ever see the light of day? - Measuring economic outcomes from very high resolution imagery

18th Apr 2017

Author: Tanner Regan

The world is routinely photographed from space on a vast scale. This abundant source of information can help answer economic questions that have been, until now, hindered by a lack of suitable data. How can economists put highly detailed daytime photographs to use? This question is only now beginning to be answered leaving tremendous potential for the future.

Increasingly, economists are turning to satellite imagery in order to fill gaps in conventional data. Donaldson and Storeygard (2016) give a nice overview of this data revolution focusing on three main advantages: revealing information that is not available conventionally, providing unusually high spectral resolution, and having wide geographic coverage.

In Henderson, Regan, and Venables (2017) we capitalize on the spatial detail of very high resolution daytime imagery to measure the volume and precise location of every building in the city of Nairobi. Focusing on buildings we are investigating a massive stock of wealth – estimated by Kunte et. al. (1998) to be 72 percent of the total value of a nation’s physical capital. Combining this data with a theoretical framework gives a novel approach to examining the little known issues regarding the evolution of the urban built environment in developing countries.

This post discusses how the building shapefile data used in Henderson, Regan, and Venables (2017) was processed in a way to make it accessible and interpretable for economic analysis, and also some alternative ways in which other researchers are making use of similar data.

Very High Resolution imagery and its derivatives

Satellite or aerial imagery is called very high resolution (VHR) when the distance between pixels corresponds to less than one metre on the ground. Such detail allows for a massive amount of information to be recorded in, say, an image of a city. To get an idea of the level of resolution, think about building a new shed in your backyard, or extending your kitchen to include a dining space, these are the kind of changes that we’re able to capture. All that the human eye can see, however, is not always easily quantifiable in an economically interpretable way. Compared to nightlights data for instance, which has a straightforward interpretation in the brightness of pixels, VHR imagery is much more complicated to untangle.

The dimension that we focus on here is the ability to trace the footprints of buildings from these images. There is work at the frontier of remote sensing research that attempts to do this using automatic and semi-automatic algorithms, including the efforts of Neeraj Baruah on the Urbanisation in Developing Countries Programme. As of yet, though, for the identification of individual buildings, only manual digitization has proved reliable enough for the economics literature. Companies like Ramani Geosystems, based in Nairobi, have teams of digitizers that go through images by hand and digitally trace building footprints.

The spectral resolution of VHR imagery also deserves some recognition. Most VHR products come either as panchromatic or with three additional bands for each of the red, green, and blue sections of the EM spectrum, although at lower spatial resolution, basically giving use a choice between black and white photos or ones in colour. However, there are some products which provide even more bands giving us an idea of the spectral signature outside of the visible light spectrum.

Building the city measures

In Henderson, Regan, and Venables (2017) we use two cross sections – one circa 2004 and another from 2015 – of building footprint shapefiles which were digitized from aerial imagery with resolution under 40cm. In order to take full advantage of the spatial detail and variation across time we developed an algorithm that determines whether a given building has been demolished and replaced or remained unchanged over the decade between cross sections. The final product is a two-period panel of individual buildings, allowing us to examine building level changes.  Figure 1 shows how the algorithm has identified demolished and redeveloped buildings in a slum neighbourhood.

With only the use of two dimensional imagery, though, we still lack the more economically relevant measure: building volume. Without accounting for height differences between buildings both a bungalow and a skyscraper could appear the same. Recognizing this, we augment our footprints with LiDAR data available in 2015 to determine the heights of individual buildings. For buildings in 2004 that were demolished, and therefore have no directly observable height data, we make a conservative approximation about heights in 2004. Using only buildings classified as unchanged by our algorithm we calculate a neighbourhood average height and assign it to the demolished buildings.

The data, with a bit of creative manipulation, allows us to track the development of building volumes across the city. Where is the built volume to land area ratio high? Which areas are changing, and if they are how much volume is added by redevelopment vs. infilling? One interesting fact that we have uncovered is that the average height of redeveloped buildings in slum areas is nearly indistinguishable from unchanged heights, while in the formal sector redeveloped buildings near the CBD are typically 50% taller than those that were unchanged.


Figure 1: Demolition in 2004 (left), and redevelopment in 2015 (right) of buildings in Kibera, Nairobi.


Other interesting applications in economics

In a paper exploring ethnic patronage in a slum, authors Marx, Stoker, and Suri (2016) make use of panchromatic imagery at 0.5 metre resolution. They take a novel approach to measuring household investment by the brightness of roofs. Noticing that roofs are almost exclusively metal sheet in their slum neighbourhoods, and that tin roofs appear much brighter when new and dull and rusted over time, they measure the average brightness of the image for each building to proxy for household investment.

Ongoing work by Guy Michaels, Jamila Nigmatulina, Neeraj Baruah, and Tanner Regan at LSE and Ferdinand Rauch at Oxford use VHR imagery and building shapefiles to create a variety of outcomes to measure the long run performance of World Bank slum upgrading and pre-emptive servicing programs in Tanzania. Using just the building shapefiles we create measures of crowding and the organisation of neighbourhood layouts. For instance a building is crowded if it has another building within one metre, and a building is unordered if the angle between itself and its nearest neighbour is far from 90 degrees. We also use multi-spectral VHR imagery in order to determine the roof material for each building. Similar to the idea in Marx, Stoker, and Suri (2016), but also allowing for non-metal roof types, we proxy household investment based on the type of roof. Using multi-spectral imagery allows us to create a spectral signature for each rooftop and see how closely this signature resembles those of tile, metal, grass, or concrete. Figure 2 shows the patterns that we try to capture, the north-west was serviced in the 1970s and is now well ordered and the roofs are speckled with colour signalling high household investment, while the south-east is unordered, crowded, and mostly rusted grey roofs.


Figure 2: Sinza (north west) and Manzese (south east) neighbourhoods in Dar es Salaam, Tanzania.