OREANDA-NEWS. March 28, 2016. Editor’s note: This is one in a series of profiles of five finalists for NVIDIA’s 2016 Global Impact Award, which provides \\$150,000 to researchers using NVIDIA technology for groundbreaking work that addresses social, humanitarian and environmental problems.

Eradicating worldwide poverty by 2030 is the top goal on the United Nations’ sustainable development agenda, published late last year. But a lack of data has frustrated efforts to measure progress toward the goal.

Most of those living in extreme poverty are in sub-Saharan Africa and Southern Asia, where accurate poverty data is scarce. A small team at Stanford University is changing that, one satellite image at a time.

Machine learning expert Stefano Ermon partnered with food security specialists David Lobell and Marshall Burke, plus a couple Stanford engineering students, to turn Google Earth images into statistical poverty models.

“We want to end extreme poverty, but we need a way to be able to measure whether we’re making progress or not,” said Ermon, an assistant professor of computer science at Stanford.

Left: Predicted poverty probabilities at a fine-grained 10km ? 10km block level. Middle: Predicted poverty probabilities aggregated at the district-level. Right: 2005 survey results for comparison (World Resources Institute 2009).
Left: Predicted poverty probabilities at a fine-grained 10km ? 10km block level. Middle: Predicted poverty probabilities aggregated at the district-level. Right: 2005 survey results for comparison (World Resources Institute 2009).

Using NVIDIA GPUs, the team trained a neural network to accurately predict poverty levels in sub-Saharan Africa from image features like roads, farmlands and homes.

This work has placed Stanford among five finalists for NVIDIA’s 2016 Global Impact Award. Each year, we award a \\$150,000 grant to researchers using NVIDIA technology for groundbreaking work that addresses social, humanitarian and environmental problems.

“There are countries in sub-Saharan Africa for which the most recent data we have is 20 years old, so we’re still extrapolating from early ‘90s estimates,” said Ermon. “There’s really a dire need for better data.”

Traditional deep learning solutions require a training dataset to build a neural network. Since there is very little training data for satellite images, the Stanford team uses transfer learning — teaching a machine skills for one task that can be transferred to a different application.

First, satellite images showing day and nighttime views of the same area are pulled in pairs from Google Earth and Google Images. Traditional poverty imaging models use the intensity of nighttime light in a photo as a measure of economic development. By using both day and night images, the model learned to identify useful daytime features correlated with poverty from the unlabeled dataset such as roads, farmland and bodies of water.

“We came up with a fully convolutional architecture that uses a smaller number of parameters and can handle larger images — and we can train it faster using NVIDIA GPUs,” said Ermon.

Left: Each row shows five maximally activating images for a different filter in the fifth convolutional layer of the CNN trained on the nighttime light intensity prediction problem. The first filter (first row) activates for urban areas. The second filter activates for farmland and grid-like patterns. The third filter activates for roads. The fourth filter activates for water, plains, and forests, terrains contributing similarly to nighttime light intensity. Right: Filter activations for the corresponding images on the left. Filters mostly activate on the relevant portions of the image. For example, in the third row, the strongest activations coincide with the road segments. Best seen in color. See the companion technical report for more visualizations (Xie et al. 2015). Images from Google Static Maps.
Left: Each row shows five maximally activating images for a different filter in neural network: urban areas; farmland and grid-like patterns; roads; water, plains, and forests. Right: Filter activations for the corresponding images on the left. Images from Google Static Maps.

The team used our GeForce GTX TITAN X and Tesla K40 GPUs to accelerate their image analysis. Training the final model took just three days on GPU-accelerated libraries.

“It was really instrumental,” said Ermon. “We could not have done it without GPUs.”

Ermon and his team recently published a paper on their research approach for poverty mapping in Uganda. The team’s entire pipeline — from downloading the images to training the network and using it to make predictions — will soon be available on GitHub.

They are now scaling up their work to satellite images from Nigeria, Malawi and Rwanda. Despite being trained on data from Uganda, the Stanford team’s system can also accurately predict poverty levels in other countries. Ermon hopes to expand the model to map poverty in Asia and South America, and to analyze changes in poverty over time.

“Now we have this platform that can automatically generate these poverty maps with very high resolution using very inexpensive data — we just need images,” he said. This scalable model opens up new possibilities for poverty mapping over larger time scales. “We can start to get unprecedented understanding of how things change over space and time.”

The winner of the 2016 Global Impact Award will be announced at the GPU Technology Conference, April 4-7, in Silicon Valley.