Covering More Ground With an Assist from Artificial Intelligence (AI)

Covering More Ground With an Assist from Artificial Intelligence (AI)

Michael Cosh is the Research Leader of the ARS Hydrology and Remote Sensing Laboratory in Beltsville, MD. His research investigates land surface information about like soil moisture and crop status from remote sensing platforms like drones, aircraft, and satellites.

Welcome, Dr. Cosh, to Under the Microscope. 

UM: Why is it helpful for farmers to have data about factors like soil moisture, evaporation, crop mapping and yield, or the presence of weeds? How do they use this information?

MC: Farmers use this information for planning and management of their operations – everything from timing of fertilizer application to how much to irrigate, and more. When it comes to understanding a complex process like growing crops, the best data is real world information. Farmers collect a new set of data each year with unique patterns of rainfall and temperature conditions, and variation in time and space. Documenting this information and building a record helps us to generate models and forecasts for the future, based on real data. 

UM: What limitations have researchers faced in collecting and analyzing this kind of data?

MC: Historically, a major challenge has been having the human resources to collect information about all the factors that are important – things like rainfall, temperature, etc. In the past, individual people had to manually collect the data, and then track those measurements and perform computations using limited computing power. Now, with AI, we have a lot of computing power available, but the big challenge is identifying, collecting, and managing all of the data that may be necessary to get the AI systems set up to function well in the first place. One way to understand AI is that it is a system of mathematical models. In order to provide useful information or other outputs, an AI system needs to develop internal parameters – guidelines that tell it what to do with information and how to assess it. To do that, it needs to start with some data to build and refine those models. We call that initial information “training data,” and it’s a vital, foundational part of the creation of any AI system. 

If you think about precipitation, for example, a mathematical model used for AI might need to know about daily precipitation. A simple rain gauge can provide that information. But it doesn't measure the amount of water in snow, which can be important for some locations (like Colorado), and less important for others (like Florida). Also, daily precipitation may not be good enough to use for some purposes – hourly or even per-minute rainfall is important, especially for soil erosion and flooding events. In addition, data on temperature, sleet, hail, and many other conditions would also be needed to better inform an AI system that is concerned with precipitation. Designing a system that captures all of this information accurately requires a lot of time, effort, and forethought about potentially valuable information, requiring subject matter experts throughout the process. In addition, locating data can be challenging; sometimes data we wish had been collected in the past cannot be recovered or reconstructed.   

A soil moisture sensor installed at a depth of 20 cm can provide long-term records of soil moisture and soil temperature,
which are invaluable for planning of planting and water availability. (Photo by Mike Cosh, ARS)

UM: How and why did researchers at ARS begin to use AI in the areas you study? What new capabilities did it offer?

MC: AI is a broad term for many different processes. One of the earliest situations where we used the technology that became known as AI was in studying land cover classification, which

is a way of describing the types of crops and other plants growing on different areas of land. That research involved using satellites to identify which plants were growing in which areas. Over the course of a growing season, USDA was able to develop a time series of images – that is, a sequence of pictures of the same areas as they changed over time. Using that time series, AI can be trained to interpret different satellite signals for other fields to reveal the types of crops growing there. That information in turn provides information for the commodity markets for those crops, and supports overall food security. 

In this case specifically, the training data were for a set of fields where we have confirmed the crop types based on human observations on the ground. We assume that the patterns we see are common for the same land cover or crop, so our small amount of human-observed data can then be applied across the U.S. So, AI provided us with a new capability in expanding the amount of land we could address; instead of being limited to a small area we had physically measured, we could extrapolate using AI to describe land conditions in many other areas as well. 

AI in Agriculture Image

UM: What role do human judgement and oversight play? Will they always be necessary?

MC: There will always be a role for humans in the process, because we will always need to verify the success and practicality of AI solutions. There are many AI applications that might seem fully mature now, but ultimately, they are probably just parts of a larger decision-making process, and that will require human input. A car may be able to drive you to a destination, but the decision to go to that destination in the first place is up to a human, and a human would have to decide whether the trip was a success or not. In a similar way, our research will always need a human to determine goals and evaluate outcomes. 

UM: Could AI ultimately replace prior approaches you’ve used in your research, or does it simply enhance them?

MC: AI is a tool first and foremost, bringing necessary information to a human for a purpose or decision to be made, hopefully in a more efficient manner. The efficiency and cost will determine the adoption of AI. If it cannot replicate the accuracy of prior approaches, it won't replace them. 

A SNOTEL, or Snow Telemetry, site in Utah provides weather and soil monitoring information on a long-term basis,
including snowpack, soil moisture, soil temperature, and other atmospheric information. (Photo by Mike Cosh, ARS)

UM: How might farmers see the effects of your use of AI in their day-to-day operations? Are there specific benefits it will deliver?

MC: There are several ways farmers might see and benefit from the AI in our laboratory. Some of our studies are looking at how AI can better estimate soil moisture at the surface, which can help farmers with irrigation scheduling, or inform ranchers who need to know how many cattle they can put into a pasture. Analysis of nutrient flows and water quality may provide better guidance on how and when to apply fertilizers to maximize long-term yield and minimize environmental impacts.

Those are just a few of the potential applications; as we continue to expand our use of AI in research, it’s likely that we will find new ways to apply it to enhance the tools we can provide to farmers and other stakeholders. In fact, right now there are already web sites that farmers and the general public can use for things like decision-making about land purchases or other issues where the characteristics of the land are important. 

UM: What are your hopes for how AI ideally might contribute to your research, and to farmers and consumers, as it is further developed?

MC: The long-term goal of AI is really to help free time for farmers and producers so they can focus on more human-centric decision making. AI is a tool that we hope will make processes more efficient and accurate, not removing farmers from the system, but helping them to be more productive and to have a better quality of life.