Farming is grueling business. Between 2013 and 2016, U.S. farmers and ranchers weathered a 45% dip in net farm income — the largest since the great depression — while the number of mouths to feed grew sharply by the day. It’s anticipated that the global population will increase by 2.2 billion by 2050, and that the world’s farmers will have to grow about 70% more food than what is now produced.
If you ask Microsoft, the solution to lies in technology. The tech giant’s FarmBeats program, which launched in preview today on Azure Marketplace ahead of Ignite 2019 in Las Vegas, is a multi-year effort to bring robust data analytics to the agriculture sector. With a backend built on Azure and compatibility with hardware from a range of top manufacturers, the goal is to promote what FarmBeats project lead and chief scientist at Azure Global Ranveer Chandra calls “data-driven” farming techniques, which the International Food Policy Research Institute claims can boost farm productivity by as much as 67% while reducing resource usage.
“[We’re capturing] large amounts of data from farm[s] and then us[ing] AI and machine learning to translate that data into insights for the growers … When we talk to growers, a lot of the decisions they make are based on guesswork, and we want to replace that guesswork with data,” Chandra told VentureBeat in a phone interview. “It’s not just about growing more food. We need to grow more nutritious food, good food, and we need to grow that increased good food without harming the environment.”
FarmBeats kicked off in 2015 with a prototype for an internet of things (IoT) platform for agriculture — a platform that enabled “seamless” data collection from sensors, cameras, and drones. Chandra drew personal inspiration from his grandparents’ farm in India, and from an Accenture survey that found that less than 20% of farmers use sensors, drones, and other tech for crop planning, owing to costs and flaky connectivity.
Within two and a half years, Chandra and colleagues had a production-ready system within Microsoft’s AI for Earth, a program that provides tools to organizations for environmental solutions development. To test its efficacy, in 2017, they seeded FarmBeats at two U.S. farms — the five-acre Dancing Crow Farm in Carnation, Washington and the 100-acre community-supported Essex Farm in Upstate New York — for six months as part of a pilot study.
Gateways and base stations
FarmBeats leverages unlicensed TV white spaces — the radio frequencies allocated to broadcasting services — to establish a high-bandwidth link from a farmer’s home internet connection to a base station, sometimes supplemented by the open source long-range IoT protocol LoRa. Sensors, drones, and the like connect to said base station, which draws power from a battery-backed solar panel pack.
The base station has three components: a TV white space transmission device, a Wi-Fi connectivity module, and a controller. The Wi-Fi module lets farmers connect off-the-shelf soil temperature, pH, carbon dioxide, and moisture sensors, as well as their phones, to access farming productivity apps. As for the controller, it’s responsible for caching collected data when the TV white space device is switched on, and for planning and enforcing power cycle rates depending on the current battery status.
“Imagine if you had a Wi-Fi router and you could access it from a few miles away … That’s what the TV white spaces enable,” said Chandra. “If you turn on a TV in the middle of a farm, most of what you’ll see is white noise. And the more noisy TV channels there are, the more unused capacity there is. Even if there are 20 TV channels that are not in use, we’re talking about close to 400Mbps [of capacity] available.”
Cost savings were a principle focus from the start. To that end, the TV white spaces radios and components supporting sensor connectivity cost around $200 and $20, respectively, and FarmBeats doesn’t have additional data charges other than the farmer’s existing internet. The end goal is to make available a full processing unit for about $80, which compares favorably with the over $1,000 price tag of conventional logging devices and base stations, and associated subscription fees exceeding $100 per month.
In the pursuit of robustness and energy efficiency, FarmBeats adopts what Chandra calls a “weather-aware” design that uses forecasts from OpenWeather and other sources to power-cycle different base station components. (In the case of Dancing Crow Farm, the team sourced seven years of data from weather stations across Washington state, which they used to train a machine learning model to make hyperlocal predictions.) FarmBeats aims for energy neutrality — that is to say, it consumes at max as much power as can be harvested from the solar panels — and to minimize data gaps, or continuous time-intervals without sensor measurements.
One of the ways it achieves this is by switching off the TV white spaces device (which consumes five times more power than the rest of the base station) at night, when it’s unlikely to be used. This dramatically decreased the amount of downtime FarmBeats faced, according to Chandra; an earlier version of the system that wasn’t power-aware exhausted its battery 30% of the time in a cloudy month.
“Solar panels and batteries don’t always work as expected,” Chandra said, “so we came up with an intelligent algorithm … What we showed is that using this system, the system could last 30 cloudy days [instead of] four.”
To mitigate connectivity problems, FarmBeats employs a gateway-based design wherein a PC at the farmer’s home performs computation locally on data, consolidating and beaming it to Microsoft’s Azure IoT Suite for long-term and cross-farm analytics. The gateway can operate independently during periods of network outage, and it runs a web service for the farmer to access detailed data within a portal — FarmBeats Data Hub — when they’re on the farm network.
Drones are a critical part of the FarmBeats system, which treats them like first-class data-gathering citizens.
Most drones retrofitted for agricultural work fly to locations in an east-to-west fashion, but Chandra and team reasoned that optimizing for area covered in a single flight was a better approach. They implemented a planning algorithm that minimizes the number of waypoints required, which in turn minimizes the time the drone spends decelerating to come to a halt before turning and accelerating again.
The team also noted the fact that farms typically have large and windy open spaces, and that these characteristics could be exploited to boost overall drone flight efficiency. When a drone directed by FarmBeats is flying downwind, a yaw control algorithm pivots the drone’s side profile (which has a larger area) to face the wind, achieving a speed boost. And right before the drone comes to a stop, the algorithm directs it to fly perpendicular to its flight path, a move analogous to a skier’s hockey stop.
The net effect: During the two-farm pilot, FarmBeats reduced the time it took to cover a given area by 26%.
The next challenge was figuring out a way to transfer drone-captured photography to the cloud. Instead of retaining the entirety of the drone’s video, the FarmBeats gateway leverages AI to estimate the position of different frames and create a geo-referenced, low-altitude panoramic overview incorporating sensor data. This is passed onto commodity stitching software, which creates a roughly 40-megapixel final image detailed enough to reveal water puddles, the location of individual cows grazing on the farm, and more.
In these images, FarmBeats infers measurements from relatively few sensors deployed on the farm, relying on the fact that similar-looking areas are likely to have the same sensor readings. (For example, recently irrigated areas with high moisture look darker.) This allows it to extrapolate and predict metrics like wetness even when no sensors are present, as well as pH levels and temperature.
“If you have a farm, and you wanted to get a [map] of soil moisture levels in the farm six inches below the soil, you would need a sensor every ten meters, but [that’s] expensive to deploy and manage,” said Chandra. “[FarmBeats is] able to reduce the number of sensors needed in the farm to actually build meaningful, actionable algorithms.”
At the two farms that participated in the pilot, Chandra and colleagues installed cameras encased in weatherproof boxes and over 100 different individual sensors. They deployed DJI Phantom 2, Phantom 3 and Inspire 1 quadcopters for aerial monitoring and created a control app using DJI’s software development kit, which allowed users to specify the flight altitude and the area (or areas) to be covered. Lastly, they procured a Raspberry Pi 3 and laptops to serve as the base station controller and gateway, respectively, which handled the transmission of sensor readings, camera images, and drone video summaries through Azure IoT Hub to storage.
Over 10 million sensor measurements, 1 million camera images, and 100 drone videos were collected over a sixth-month period.
Farmers at the New York location took advantage of FarmBeats’ extensibility to monitor temperature-regulated storage freezers, and to trigger mobile alerts when the temperature dipped below a certain threshold. They also plugged in cameras at places like cow sheds and connected them to the base station, which the FarmBeats team experimentally used with an AI model to identify cows as they came within view.
Microsoft subsequently teamed up with Nelson farm in Spokane, Washington to conduct an extended trial of the FarmBeats system. Over the course of 11 months, they report that it helped to reduce the application of expensive chemicals by 90%, increasing cost savings increased by 15%. Moreover, the system’s temperature predictions enabled the farmer to take preventative measures ahead of deep freezes (dips in temperature below freezing within 24 hours).
FarmBeats has since been implemented in farms across the U.S., as well as India, New Zealand, and Kenya, and soon, Brazil. Simultaneously, Microsoft has inked agreements with drone and sensor manufacturers and weather and satellite imagery providers like Davis Instruments, DJI, SenseFly, Sentinel, and the European Space Agency to furnish FarmBeats users with equipment and data. Through its Solution Accelerator program, Microsoft will license tools out to partners for device management, like machine learning models that recommend the number of sensors to cover a given farm.
In March, Microsoft collaborated with the National FFA Organization to deliver 100 FarmBeats Student Kits to 50 chapters. The kits include a Raspberry Pi computer with soil moisture, light, ambient temperature, and humidity sensors. Last month, Microsoft and the USDA jointly announced FarmBeats’ latest expansion: a 7,000-acre farm at the agency’s Beltsville Agricultural Research Center in Maryland that will pilot the system for cover crops, or crops grown during the off-season to limit weeds and pests and improve soil.
The USDA will roll out FarmBeats to the 200-plus farms in its research network if all goes well, which USDA research ecologist Steven B. Mirsky says could enormously benefit the Agricultural Research Service. The agency’s principle in-house research arm employs roughly 2,000 scientists across 90 research locations and regularly collects and analyzes field sensor data for ongoing research.
“One of the things we struggle with is trying to figure out how to get that information out of … data silos,” Mirsky told VentureBeat in a phone interview. “Typically, within our research units, data … stays housed within those locations until they do their publications. Our goal is to get as much of the data from [those] research locations as we can back up into [our] cloud-based data repository. [With FarmBeats,] we can really add additional power to our models and our decision support tools.”
It’s key to note that for all of Microsoft’s philanthropic overtures, FarmBeats is partly a play for the estimated $5 trillion agriculture market. It has competition from a number of companies actively building data analytics solutions for farmers, including North Carolina startup Growers, Understory, Trace Genomics, EarthSense, Solinftec, TeleSense, Pollen Systems, and industry giants such as John Deere and Monsanto.
But FarmBeats is only the start for Microsoft. Work is underway on Project Sonoma, a project that aims to create a commercial-ready autonomous greenhouse for vegetable production.
“[With all this data,] you need the ability to combine all of these data streams,” said Chandra. “That’s what Azure provides. Our goal is to make it the best cloud for any agricultural application.”