Droughts and Deep Learning: Measuring Water Where It’s Scarce

by Jamie Beckett

Withering droughts that recently parched areas as far-flung as California, Brazil and Australia are just a preview of what’s to come, scientists predict, as global warming worsens.

We can’t make it rain. But deep learning could help thirsty communities find out just how much water they have.

Orbital Insight, a Palo Alto, Calif.,-based startup, has developed a GPU-accelerated deep learning system that’s designed to measure and monitor the surface levels of freshwater reserves across the globe.

Using surface area to indicate water volume, the company aims to start giving water managers and policymakers a picture of how much water is available.

“To understand the severity of a drought, you have to measure its effects,” said Shwetank Kumar, the company’s vice president of product engineering.

Once communities know how much water is available, they will be able to make better decisions about how to use and save water, he said.

Deep Learning Detects Water 

The process starts with images from the U.S. Geological Survey’s (USGS) Landsat 8 satellite, which images the entire Earth every 16 days. Orbital Insight feeds these into a neural network that searches for water, pixel by pixel.

Droughts and deep learning: Record-low water levels in California's Folsom Lake during drought.
Record-low water levels in California’s Folsom Lake during drought. Image courtesy of Vince Migliore.

The company trained its neural network on billions of pixels contained in thousands of USGS satellite images. NVIDIA GPUs hosted in the Amazon Web Services cloud accelerated the training process.

Orbital Insight isn’t the first to measure Earth’s water levels, but its deep learning-based architecture results in higher accuracy than many conventional techniques, which often mistake shadows of clouds or mountains for water, Kumar said

“Our goal is to produce a fully automated water monitor that yields weekly updates and provides statistically significant measures of changes in water surface area,” Kumar has written.

The company tested its deep learning water measurements on reservoirs in drought-stricken California and Iraq, and is preparing to soon roll out the technology to customers.

Global Warming to Worsen Droughts 

The stakes are high. More than a billion people now live in water-scarce regions, and as many as 3.5 billion could experience water scarcity by 2025, according to the World Resources Institute, a nonprofit research organization.

The world’s arid and semi-arid regions are likely to become drier, hotter and experience longer and more frequent droughts, according to the Intergovernmental Panel on Climate Change, an international group that assesses climate change science.

But climate change is only part of the picture. Population increases, shifting diets and agricultural expansion that increase demand for water often play just as large a role in water stress, said Rutger Hofste, a research analyst for the World Resources Institute’s water team.

“Water stress is increasing in a majority of the world, especially in already vulnerable areas such as the Middle East, North Africa and Australia,” he said. When there’s not enough water, farmers often suffer and people can lose reliable access to the food they need, Hofste added.

Kumar said his company’s technology could also be used to enforce water agreements among nations or states.

“All that starts with knowing how much water is available,” he said.

water stress projections for 2040 by the World Resources Institute
Rising global temperatures, reduced precipitation in arid regions and growing demand for water will increase water stress worldwide, according to projections by the World Resources Institute. See http://bit.ly/1KxbEXf for more information about WRI’s Aqueduct projection tool.