Listen to this article

Thomas Jefferson once commented on the beauty of Lake George, N.Y., claiming the lake “is, without comparison, the most beautiful water I ever saw.” More than 225 years later, researchers are using smart sensors, computational modeling, and data analytics to make sure the water remains that way.

A collaboration among IBM Research, Rensselaer Polytechnic Institute (RPI), and The Fund for Lake George, known as The Jefferson Project, has been monitoring the lake and its surrounding ecosystem for the past five years. The project is a continuation of a 30-year study that began in 1980 with manual data measurements used to track the lake’s water quality and other environmental stressors.

“Information is power, and the Jefferson Project is providing empowered information in our concerted work to make Lake George the world’s smartest and best protected lake,” said Eric Siy, executive director of The Fund for Lake George. “Not only does this unparalleled endeavor reveal new solutions to intensifying problems facing Lake George and freshwater ecosystems anywhere, the project’s Smart Sensor Web measures our results. By doing so, it provides a positive feedback system that will ensure ongoing commitment to enduring success.”

The smart sensor network of sensor platforms, weather stations, tributary monitors, and acoustic Doppler current profiles provides data at more than 40 locations on and around the 32-mile lake in Upstate N.Y.

The network continually monitors conditions such as weather, water movement, and water quality, particularly levels of salt and oxygen content.

The collected data is then fed to an IBM Research center in Yorktown, N.Y., which creates high-resolution computational modeling and forecasting models.

Smarter sensors and updated computer modeling

Researchers on the project added computational intelligence to the sensors to make them smarter than traditional “dumb” sensors, which would only collect data and do nothing else, said Eli Dow, senior software engineer for the Jefferson Project at IBM Research.

Using the computer models and forecasts created from the real-time data, these sensors can react to the physical data and the models to collect more data or alert project members.

For example, a tributary sensor that was measuring the flow of a stream feeding water into the lake would have information based on what the last few observations were (older data), plus what the next likely observations would be (forecast data). If an anomaly then occurred, such as the water rising faster than expected, the sensor is intelligent enough to alert officials about the anomaly, Dow said.

“So as we’re measuring whatever values these sensors are recording, they can look for anomalies on the fly and say, ‘Something looks suspicious, so maybe we should measure more frequently and figure out what this is,’” Dow said. “Whether it’s something like, ‘The stream will rise really quickly because of a snow melt event, or there’s lots of rain coming in the area,’ we can respond and react based on the measurement stream coming in – the anomalies – or based on proactive forecasts.”

To get the level of data needed for the computational modeling and forecasting, a larger sensor network was required. Mike Kelly, a senior research engineer at IBM Research, said the original 30-year study, which used manual physical sampling once or twice per month, was able to identify some long-term stressors on the lake, but that more frequent sampling would be required.

“There was an assumption that measuring the system at that frequency, while it would reveal some long-term trends, was insufficient to really understand the dynamics of the lake,” Kelly said. “Those dynamics within the lake and the watershed could be very important to understand how these trends are affecting the lake and how they could be reversed.”

The smart sensor network includes several different types of environmental sensors, including:

• Weather stations, which provide data on wind speed, wind direction, air temperature, humidity, pressure, precipitation, and solar irradiance.
• Water movement sensors, which measure tributary water flow and level, 3D volumetric water speed, and 3D water direction.
• Water quality sensors, which measure conductivity, water temperature, pH levels, dissolved oxygen, chlorophyll levels, dissolved organic matter, and chloride levels.

At the five deepest points in the lake, researchers have placed sensor platforms with vertical profilers. This includes a semi-robotic winch system that sends sensors down through the water column, taking measurements along the way from the surface to the lake bed, Kelly said. While the system is not complex from a mechanical point of view, operating the systems in harsh environments can be tricky.

“These platforms, even though they’re moored out on the lake, can be subjected to four-foot waves at pretty high frequency,” Kelly said. “So these platforms can be bouncing around quite vigorously, all the while you’re trying to position this sensor package at a specific depth in the water column. So it’s very challenging from both a hardware and software perspective.”

With the sensor network operational and data measured, researchers then let the computational models take over.

On the weather side, Dow said the project is using IBM Research’s Deep Thunder tool to get hyperlocal forecasts every 10 minutes.

“The weather model is foundational in that it drives the other models,” Dow said. “With a narrow lake such as Lake George, a standard weather model may operate at a relatively coarse resolution of several kilometers, which is insufficient to represent or capture the complexities of the lake and surrounding watershed.”

Through the research, the project created a weather model that currently runs at a 333-meter (1,092-ft.) resolution, giving them forecast points every 333 meters for the entire watershed and lake. The models also make predictions about water runoff to determine where rain or snow goes into the soil or other streams, eventually entering the lake.

This intelligence can give better insights to decision makers on processes and events that would, or could, affect the lake’s ecosystem.

Dow gave another example – a weather forecast made in the middle of February, when the ground is frozen, that says a snowstorm or freezing rain is coming. The town department of public works decides to presalt the roads ahead of the storm, but then the weather changes to rain, causing the salt to flow into the streams that feed the lake.

With more accurate forecasting via the computer models, the system could inform the local department of public works to not salt those areas as much because of existing high salt content.

Moving forward with the new tech, new models

The team said it continues to apply new technologies to address problems and to better collect and analyze data, both at the multi-sensor platform and at information platform level.

On the hardware side, the team is considering underwater autonomous vehicles and other new technologies that can be connected to the smart sensor network.

Beyond the measurements of the lake’s ecosystem — see box below — the team is able to use what it has learned for other environments and business problems.

“The way I look at it is Lake George and the Jefferson Project really remains our laboratory, not only from the environmental perspective of the lake itself, but [also] the infrastructure of the smart sensor network, the operational modeling and the ongoing scientific research,” said Kelly.

With the infrastructure in place, he added, “that allows us to expand our capabilities of how we might map this technology to another area, or to a whole different type of environment.”

On the computer modeling side, other projects don’t even need to have a water-based problem, Dow said.

“From one vantage point, it’s about big data and making sense of that big data,” he said. “The machine-learning analytics endpoint software and hardware that we’ve co-developed — there really isn’t anything inherently water-based about that. It just takes signals in from sensors and can do these kind of reactions based on what it’s learned about its own environment locally, or what it gathers through forecasts that have been sent down from a cluster drive model.”