The Artificial Intelligence revolution is upon us, and agriculture will see wide-ranging effects. Many of agriculture’s hardest problems are being solved by companies with the right combination of data and the technical expertise in Artificial Intelligence (AI) techniques.
Machine learning, which can be considered a subset of the broader trend in AI, has arisen from the confluence of several long-term trends in technology, most notably the proliferation of real-time measurements of physical systems (i.e., The Internet of Things) and the decreasing costs of cloud computing.
The story of the development of a AI-powered irrigation product by our company, Tule, highlights the unique opportunities and challenges for machine learning solutions in agriculture.
IoT + Cloud: sensors in the physical world transmit data to the cloud for processing and analysis. Machine learning systems enabled by powerful cloud computing systems produce insights accessible from computers and smartphones.
The Internet of Things for evapotranspiration. This map shows Tule’s sensor network in one area of California. High density of sensors enables collection of measured ET data at unprecedented scale.
An eddy covariance system, used for measuring evapotranspiration in academic research, is too complicated and expensive for growers to use.
Tule’s sensor, which measures evapotranspiration with the surface renewal method, is used in large scale production agriculture by hundreds of growers.
The Challenge For Irrigation Managers and the Opportunity for Machine Learning
Understanding the current water status of a field and predicting how it will respond to irrigation are problems that growers face every time they make an irrigation decision. Even veteran irrigation managers admit that the process is one of guesswork, and their guesses often miss the mark.
A moisture release curve describes the relationship between available water in the soil and plant water status. Illustration from Dr. Ted Hsiao’s text on Plant-Water Relations.
When water availability decreases, plants experience water stress. Corn leaf rolling is an example of a physiological response to water stress.
Tule’s site-specific MRC model shows growers how long until a field drops below the grower’s desired level of crop stress and provides irrigation recommendations that will keep a field within the desired stress range.
Each farm field is a unique and complex biological system. Farm management operation decisions, such as fertilizer and irrigation applications, require careful consideration of the many interconnected pieces of agroecological information about each field. Therefore, farming decisions are ideal candidates for the application of machine learning, which excels at making sense of vast and complicated datasets.
Let’s take irrigation decisions as an example. When irrigation managers talk about how they make irrigation decisions for a particular field, their method usually involves a combination of knowledge about the type of crop, information about the unique attributes of the field, and recent impressions about the field collected by their scouts. In their minds, they are building a mental model of each field that helps them understand the relationship between irrigation and the crop stress of the field. In plant science, that relationship is called a Moisture Release Curve.
Leveraging its expansive dataset on daily water inputs (precipitation and irrigation), water outputs (measured ET), and crop stress, Tule developed a machine-learned Moisture Release Curve. (Note: Measured ET is the amount of water a crop field uses, whereas Reference ET, or ETo, which is a more familiar term to many growers, is an estimate of the water use of a well-watered lawn.) Like any machine learning model for agriculture, it had to meet the following criteria:
- Ability to generalize: the model needs to be sufficiently complex to successfully capture the heterogeneity that is inherent to farming.
- Speed of convergence: the model needs to provide useful predictions after a relatively short amount of time. A model that needs to be trained on an entire season’s worth of data for each field won’t be useful in practice.
- Explainability: In agriculture, it is important to be able to provide not just a prediction, but also to explain the “why” behind the prediction in order to help growers trust the data enough to act on it.
Tule’s AI-based Moisture Release Curve feature, called Water Stress Forecasts, gives growers the ability to quickly and easily see the influence of different irrigation timings and amounts on the future plant water status of their fields.
The Future of AI for Agriculture: More collaboration and more solutions
There are many opportunities for AI solutions in agriculture, but they require bringing together datasets currently housed in the separate silos of different agtech companies. To unlock solutions to these problems, agtech companies will need to partner to enable collaboration. Tule’s recently announced API partnership program is one example of this trend in action. You can read more details about the API partnership program and the development of Tule’s AI irrigation algorithm at:
https://tuletechnologies.com/tule-stress-forecast-case-study