Can new developments in machine learning and satellite imagery be used to estimate jobs?
By Alvaro Gonzalez
Methods that use satellite data and machine learning present a good peek into how Big Data and new analytical methods will change how we measure poverty. I am not a poverty specialist, so I am wondering if these data and techniques can help in how we estimate job growth.
Researchers use machine learning — using algorithms to learn from data — combined with satellite imagery to predict the distribution of poverty. Combining a time series of these images with computers that learn to recognize buildings, roads, vehicles and signs of economic activity, yields increasingly more accurate, less expensive, and more scalable methods for estimating consumption expenditure and asset wealth. Income growth in pictures.
To accomplish the same feat without satellites and computers, you need an army of enumerators carrying clipboards and pencils out in the field, interviewing people about their changing fortunes. It also costs a fortune. It is also slow and tedious work. As a final point, surveying is prone to miss the folks we really care about like the poor located far from infrastructure like roads and electricity.
In estimating job growth, we face the same kind of problems: Labor force surveys keep us informed of the job market (from the supply side), but they are expensive and time consuming. But what about firms (the demand side of the labor market). Business censuses or surveys, if they exist, are more sporadic than rain in the Atacama Desert so we cannot depend on these to keep track of jobs growth either. The natural inclination is to ponder whether the same techniques being developed to estimate poverty, with some tweaks, have the promise to deliver as well for measuring jobs created.
The “tweaks” required to estimate job growth to the emerging machine learning techniques is a modest way of saying that the challenges are substantial but likely surmountable. When the goal is estimating poverty, computers are tasked to discern features of the landscape such as urban areas, roads, bodies of water, and agricultural areas. Thatched or zinc roofs on houses, for example, are also distinguishable features that machines learn to identify and all of these features are likely to be correlated with economic outcomes such as poverty — from thatched to zinc, means a household is likely better off. In sum, machines learn which features are useful for estimating economic outcomes.
The models being developed are doing increasingly well at predicting assets held by households and on consumption expenditures. From the perspective of jobs, these two variables are the outcomes from productive economic activity (formal or informal work). So, our expectation is to be able to back out jobs from these data as well. Now, getting estimates (on jobs) from estimates (on assets) is potentially piling on error upon error. The hope is to be able to get jobs estimates directly from the satellite pictures themselves; just like the machines did for assets and consumption expenditures. This is likely to be a lot harder, however, since from buildings, roads, traffic and agricultural landscapes, machines will have to be able to distinguish the potential size of the economic activity taking place and then back out how many people are employed in these. In a building that 10 by 10 meters, labor intensive or more capital intensive production could be taking place, employing hundreds of workers or just ten. Hard to know. However, do not underestimate what these machines can learn. From traffic patterns, size of vehicles entering and leaving premises, destination of the products, and other more dynamic features, these learning machines may soon be able to guess what and how much is being produced. To estimate jobs from that amalgam of information may be a more straightforward task.
There are many more challenges, but even more promise. If we could get this done, we would be able to assess, in near real time, the effect of interventions on jobs in conflict areas, for the informal sector and in remote areas. Couple these estimates with poverty maps or maps where certain firms and sectors are prevalent and we would be able to tell the tide of job creation is rising in areas in most need and sectors which employ a lot of workers. Finally, in times of economic downturn, we may be able to detect when some are feeling those effects in terms of jobs while others are about to get hit.
Because of this promise, we aim to engage with donors, other multilaterals and the development community to help us develop these tools. The Sustainable Development Goals, for one, require us to do so.
Follow at World Bank Jobs Group on Twitter @wbg_jobs.
Read more World Bank blogs.
