Take home messages

• AI offers potential improvement for broadacre agriculture but might not be the panacea that is described.
• Growers and agronomists are encouraged to engage with researchers to find opportunities for AI to deliver enhanced decision-making.
• AI-enabled agriculture requires a digital environment that needs building.

Background

Artificial Intelligence (AI) interest and speculation seem to be ever-growing across all industries. We can say that it all started with a famous publication called ‘Computing machinery and intelligence’ in the 1950s (Turing 1950). In the same decade, the building blocks of modern AI were proposed – a paper from 1958 (Rosenblatt 1958) describes the first neural network, the bricks in ChatGPT-like models' foundations. Although the first mathematical concepts for modern AI started in the 1950s, AI only really took off once computers could handle the incredible amount of math behind these models. If you were an early adopter of emails, you might have noticed that the ‘junk’ or ‘spam’ box didn’t exist, and around the 2000s, it was introduced based on AI algorithms. That was already an AI application. The first AI face recognition in social media was in place in 2015 and used in Canberra’s airport in 2017. So, AI has been part of our lives for some time now. Perhaps more recently, we have noticed that our phones are capable of transcribing what we say into text messages, and we ask our gadgets about the weather forecast for the day, and they comment that we might need a jumper. Our routes are optimised for the fastest, quickest, or most fuel-efficient paths. In our work life on the farm, we see that some boom sprayers can detect green patches in the dirt and spray accordingly, and our tractors have been guiding themselves for some time now. More recently, we’ve seen incredible achievements in generative AI models that quickly generate texts, images, sounds, and videos, such as ChatGPT. We are now surrounded by AI-enabled or empowered technology. Here, we will open the bonnet and demystify AI a bit. We will see how some of these algorithm categories are applied to agriculture – this should help those seeking to understand AI better. Then, we will venture into AI-enabled agriculture's potential prospects and perspectives.

Demystifying AI

AI is a general term that encompasses various methods and algorithms. Regardless of the AI algorithm, there are four general types of methods: classification, regression, clustering, and dimensionality reduction. They are grouped into supervised learning (classification and regression) and unsupervised learning (clustering and dimensionality reduction). Still, the four splits are easier to link to reality because they are separated based on the types of problems.

• Classification problems are those where the objective is to separate the data into classes. In more traditional statistics terminology, the output variable, in the end, is categorical or qualitative. The classic example is where the AI identifies a picture as a dog or a cat. In agriculture, classification algorithms are commonly used to identify crops from earth observation data (Paludo et al. 2020), crop disease (Rangarajan Aravind et al. 2020), to count fruits (Koirala et al. 2019), and identify paddock boundaries (Waldner and Diakogiannis 2020), to name a few examples.
• Regression problems aim to predict a number in the end, that is, a numerical or continuous variable. In agriculture, regression algorithms have been used to estimate crop yields (Richetti et al. 2018), nitrogen fertilisation rates (Colaço et al. 2024, 2021), or the impact of crop rotation on wheat yields (Lawes et al. 2022), to name a few examples.
• Clustering problems require data to be grouped based on their resemblances or contrasts. In agriculture, a common example is delineating homogeneous zones in precision agriculture (Córdoba et al. 2016), another example is by clustering agriculture management (Taşdemir and Wirnhardt 2012).
• Dimensionality reduction solves the number of variables or features in a problem. It reduces the size of the data to a more manageable size, often to be further analysed. For example, on multi-temporal satellite imagery data for crop classification (Gilbertson and van Niekerk 2017) and is often used as part of breeding analysis.

Note that the type of problem distinguishes the categories. Hence, the first task before applying any AI method is to define the problem. Understanding the categories aids in delineating the techniques and AI algorithms to be applied, plural, because a stack of multiple algorithms to solve the problem may be needed. Generative AI models, such as ChatGPT or MidJourney, that can produce text, images, sounds, or videos based on a prompt are unsupervised learning models that produce more outputs that mimic the data based on probability distributions. At heart, these models are a clever arrangement of regression neural networks with embedding and other methods to convert those numbers into the desired output.

Perspective

Imagine a fully automated farm where the decisions are aided by AI models or done by them, where robots collect the necessary data and conduct the needed operations of sowing and crop monitoring. Weed, pest, and disease controls are done with precision only where it is required. The crop is harvested, stored and traded in the best markets for the best profit. Environmental metrics are assessed and optimised for green operations, and sustainable farming practices are implemented without profit decline. Growers retain the power to choose the best path forward from various scenarios built from their own farm reality. Acquiring new land or machinery is assisted considering the nuances of each farm and updated market options. Succession planning is smoothed with facilitated formal planning. Bookkeeping and administrative tasks are readily done when required and kept secure but easily accessible. This or its most desired components will only be possible with an integrated AI system on the farm.

As an example, to develop and deploy an automated weed control system. Data from the various weeds in different stages of the crop is needed to build an AI model that can accurately identify and quantify the weeds. Once such a model is built, the monitoring data is required so the model can flag the weeds in the paddock and assemble an application map with the different options of chemicals. This needs a data ecosystem where monitoring data is available to be input into the model that will flag the need for a spraying operation with a suggested herbicide application map. There needs to be data consistency between the way the scouting for weeds happens and the development of the AI model. This can potentially augment growers' and agronomists' scouting capability and reduce costs by applying only when and where it is required.

While some components of it are being actively worked on with varying degrees of readiness, others are yet to be uncovered. For example, the decision-making on machinery upgrades and replacement schedules. When selecting boom spray, one could save hundreds of thousands of dollars by having access to relevant information with adequate analytics. This is a problem with which an AI solution could help, a current GRDC project on automation (CSP2405-022RTX) could provide further insights. Another potential for AI is with spraying operations. Identifying diseases and pests to allow localised control could save hundreds of thousands of dollars per year in farm operations, but only if the AI solution is a package that produces a timely and actionable outcome, like an application map. Another example from GRDC's investments is leveraging farm data with On-farm Experimentation for mid-season N decisions in cereals with AI (9176493). This means that AI has the potential to enhance our decision-making in agriculture, particularly in the broadacre context. The key is to focus on problems that consider all aspects of farming. That is, AI systems need to integrate both agronomic and business models. For AI systems to be effective, they need to operate in a data ecosystem. This is the first challenge: have consistent and representative data for a database that grows with time. It also means determining the key data to be collected and making such data collection effortless and part of the operations, which is important to deliver a profitable AI solution. Otherwise, why bother? This leads to an understanding that in your business, these powerful analytical tools should be deployed to assist decisions that have a big impact on profitability first and less so on basic agronomy decisions that have little impact on the profit and loss statement.

If you're a grower, start by addressing one problem and gathering relevant data. Yield maps are a good starting point, as AI can use them to define zones and identify consistently underperforming parts of the paddock, enabling targeted amelioration operations that impact the next season. AgTech and service providers play a crucial role in facilitating data capture with clear goals in mind, ensuring that the data leads to actionable outcomes. A seamless user experience is essential to avoid adding to growers' workloads. Impactful technology will be readily adopted. Researchers should not underestimate the amount of work needed to gather and organise the data; often, the model development ends up being a smaller component of the data pipeline.

The way big corporations will leverage their databases with AI is in its infancy. An improved interface with voice prompts could augment growers' and agronomists' capability to tackle a problem, providing insights and accelerating the response time when required. For example, enhancing the quickness of response to a frost event. Nonetheless, these corporations must overcome the trust issue around the perception that they might be trying to increase sales more than help growers and their advisers. These trust issues will also permeate data sharing and privacy concerns when data is used to build AI systems in agriculture.

Conclusion

AI has shown results in many research environments and some real farm applications. While investments in developing the technology are needed, the cost-benefit for some applications has been realised. Further interaction between researchers, industry partners, and growers will highlight the critical problems for successful AI applications.

Acknowledgements

The research undertaken as part of this is made possible by the significant contributions of growers through the support of the GRDC, the authors would like to thank them for their continued support.

References

Colaço AF, Richetti J, Bramley RGV, Lawes RA (2021) How will the next-generation of sensor-based decision systems look in the context of intelligent agriculture? A case-study. Field Crops Research270, 108205. https://doi.org/10.1016/j.fcr.2021.108205

Colaço AF, Whelan BM, Bramley RGV, Richetti J, Fajardo M, McCarthy AC, Perry EM, Bender A, Leo S, Fitzgerald GJ, Lawes RA (2024) Digital strategies for nitrogen management in grain production systems: lessons from multi-method assessment using on-farm experimentation. Precision Agriculture 25, 983–1013. https://doi.org/10.1007/s11119-023-10102-z

Córdoba MA, Bruno CI, Costa JL, Peralta NR, Balzarini MG (2016) Protocol for multivariate homogeneous zone delineation in precision agriculture. Biosystems Engineering 143, 95–107. https://doi.org/10.1016/j.biosystemseng.2015.12.008

Gilbertson JK, van Niekerk A (2017) Value of dimensionality reduction for crop differentiation with multi-temporal imagery and machine learning. Computers and Electronics in Agriculture 142 (Part A), 50–58. https://doi.org/10.1016/j.compag.2017.08.024

Koirala A, Walsh KB, Wang Z, McCarthy C (2019) Deep learning – method overview and review of use for fruit detection and yield estimation. Computers and Electronics in Agriculture 162, 219–234. https://doi.org/10.1016/j.compag.2019.04.017

Lawes R, Mata G, Richetti J, Fletcher A, Herrmann C (2022) Using remote sensing, process-based crop models, and machine learning to evaluate crop rotations across 20 million hectares in Western Australia. Agronomy for Sustainable Development42, 120. https://doi.org/10.1007/s13593-022-00851-y

Paludo A, Becker WR, Richetti J, Silva LC de A, Johann JA (2020) Mapping summer soybean and corn with remote sensing on Google Earth Engine cloud computing in Parana state – Brazil. International Journal of Digital Earth 13, 1624–1636. https://doi.org/10.1080/17538947.2020.1772893

Rangarajan Aravind K, Maheswari P, Raja P, Szczepański C (2020) Chapter nine – Crop disease classification using deep learning approach: an overview and a case study. In 'Deep learning for data analytics'.(Eds H. Das, C. Pradhan, N. Dey) pp. 173–195. (Academic Press) https://doi.org/10.1016/B978-0-12-819764-6.00010-7

Richetti J, Judge J, Boote KJ, Johann JA, Uribe-Opazo MA, Becker WR, Paludo A, Silva LC de A (2018) Using phenology-based enhanced vegetation index and machine learning for soybean yield estimation in Paraná State, Brazil. Journal of Applied Remote Sensing 12, 026029. https://doi.org/10.1117/1.JRS.12.026029

Rosenblatt F (1958) The perceptron: a probabilistic model for information storage and organization in the brain. Psychological Review 65, 386–408. https://doi.org/10.1037/h0042519

Taşdemir K, Wirnhardt C (2012) Neural network-based clustering for agriculture management. EURASIP Journal on Advances in Signal Processing 2012, 200. https://doi.org/10.1186/1687-6180-2012-200

Turing AM (1950) I. – Computing machinery and intelligence. Mind LIX, 433–460. https://doi.org/10.1093/mind/LIX.236.433

Waldner F, Diakogiannis FI (2020) Deep learning on edge: extracting field boundaries from satellite images with a convolutional neural network. Remote Sensing of Environment 245, 111741. https://doi.org/10.1016/j.rse.2020.111741

Contact details

Jonathan Richetti
CSIRO
147 Underwood Ave, Floreat WA 6015
0411 150 425
jonathan.richetti@csiro.com