Artificial intelligence (AI), machine learning and deep learning: what can they do for agriculture?

By Drs. Andrew Moore, Dave Henry, Rose Roche and Ashfaqur RahmanFebruary 11th, 2022

Artificial intelligence and its sub-branches machine learning and deep learning are making their way into agriculture and agribusiness. They're helping make decisions and predictions once never dreamed of.

Whether we know it or not, we all use artificial intelligence, or AI. Most of us use AI through our smart phones and other devices. For example, our smart phones track our steps and auto connect to fitness and health apps. And texting uses higher powered types of AI like machine learning and ‘deep learning’, an even newer tool.

These techniques have started to make their way into agriculture.

But can they help us farm?

What is artificial intelligence, or AI?

When it comes to farming, artificial intelligence is a fancy name for ‘turning data into useful insights’.

Dr Dave Henry, a research leader in digital agriculture at CSIRO, explains.

“AI can give us new information sooner for traditional farming decisions. But it can also give us information on new things we would once never have dreamed of,” Dr Henry explained.

“But the most exciting part of bringing AI into farming is its power to make predictions. For example, we can make an accurate, evidence-based yield prediction even before we plant a crop.”

But it’s not just AI that’s helping farmers. Improved connectivity, the Internet of Things, satellites and smart sensors are all new ways of providing farming and agribusiness decision support tools on soils, climate, livestock, crops and weather forecasts.

What is machine learning?

The branch of AI known as machine learning (often abbreviated to AI/ML) is where a computer ‘learns’ something from past data you give it in order to make predictions about the future.

An example in agriculture is the eGrazor collar system for cattle. CSIRO developed eGrazor with the NSW Department of Primary Industries. An Australian first, eGrazor uses sensors and machine learning to measure how much grass cattle are eating based on the way they move and, over time, the system builds up this data to predict pasture intake. The benefits are numerous. Knowing pasture intake is the basis for identifying the most efficient individual grazing animals, helping breed more efficient animals and improving livestock productivity and rangeland management.

A cow wearing a collar

A prototype eGrazor collar, coupled with machine learning (a branch of artificial intelligence), can help increase livestock productivity and manage rangelands.

Another example is CSIRO’s WaterWise technology.

WaterWise is the first water-use efficiency product for high value irrigated crops that measures crop water stress and predicts future water needs in real time.

Dr Rose Roche leads the project.

“Predicting water needs in real-time is complex. This predictive power is the real scientific breakthrough that we can only get from advancements in data science, AI and computer grunt,” Dr Roche said.

“It has meant we can develop real-time predictions that can take into account the variations in a field that occur over a day.”

Aquaculture is another agricultural sector that could see benefits from AI and machine learning. The company Aquabyte is developing computer vision and machine learning-based technologies for efficient stock tracking, identifying skin conditions and feed management in salmon pens.

At CSIRO, we’re researching machine learning, computer vision and augmented reality to predict catastrophic drops in water quality in prawn ponds. We’re also researching a low-cost, smart camera device to monitor the landscape for pest and wildlife management, with feral pigs as the first application, using machine learning.

A scientist with a sensor in a field.

CSIRO’s Dr Rose Roche says only with machine learning could we make real-time predictions of crop water needs.

What is deep learning?

Deep learning (DL) is a newer branch of machine learning that deals with even more complex problems. It is capable of using context to predict solutions in complex environments. For example by looking at what’s ‘nearby’, not just ‘what’s there’.

To date, there are only a few visual recognition systems globally that are currently used in agriculture. The most frequently reported systems are fruit harvesting and the identification and control of weeds.

CSIRO’s first agriculture-related deep learning application is ePaddocks™. It uses satellites to identify the boundary of every individual paddock in Australia’s grain growing region (there are around 1.7 million of them). Using satellite technology and DL algorithms, it saves farmers having to manually outline their paddocks in each agtech application they use.

A satellite image

CSIRO’s first agriculture-related deep learning application, ePaddocks™. Crops (the red area on the left), rivers (in the centre area), bushland (black) and less vegetated areas (green on the right) with paddocks lines overlaid in 74×54 kms of Western Australia. Credit: Copernicus Sentinel data

CSIRO is also working on a deep learning model, ‘HairNet’, that uses visual recognition to assess the ‘hairiness’ of cotton leaves, a key trait in the insect resistance, fibre yield and fibre value of cotton.

We have several other AI tools at various stages of development, some in conjunction with industry partners. Tools include identification of crop types by satellite (called CropID), forecasting crop yield at any scale across Australia (Graincast™, with Digital Agriculture Services) and smart ear tags for livestock (Ceres Tag).

Another example from Australia is with sugarcane in far north Queensland. James Cook University is working on a targeted robot sprayer for sugarcane that uses deep learning to detect and only spray priority weeds. The technology aims to help reduce herbicide runoff into waterways and Great Barrier Reef ecosystems.

How do farmers know they can trust AI-based tools?

In the future, more and more AI, machine learning and deep learning applications will come online in the agricultural sector. They will solve questions we never thought we could answer, including some questions we never thought to even ask!

But how can farmers know they can trust what AI-based tools are suggesting they do, especially if they challenge their gut-feelings or traditional practices?

Dr Ashfaqur Rahman leads the sensor data analytics team with Data 61, CSIRO’s digital powerhouse.

“Trust in AI/ML systems and models is a key issue. Ag industry practitioners are often reluctant to fully trust these systems. Why? Because they are considered a ‘black box’,” Dr Rahman said.

“In decision making systems, users want to know how and why a decision is made, alongside what decision is made. But machine learning is complex and can be really hard to explain. So researchers are currently looking into ‘explainable AI’ to deal with this issue.”

The right answers to the right questions

How do you know when a tool is using AI of one form or another? The short answer is that you don’t. Most technology users don’t want to see ‘the wheels turn’. We don’t want to see how long it takes to get a result from the app or tool we’re using. We just want to see the result – fast and simple.

Dr Andrew Moore is a research leader in digital agriculture at CSIRO.

“As scientists, we always need to remind ourselves to ask what’s in it for farmers, businesses or communities?” Dr Moore said.

A phone with graphs showing water quality

One of our AI tools for the aquaculture industry that predicts dissolved oxygen in prawn ponds. This is a key requirement for water quality and prawn health.

“We can build an app or tool that gives all sorts of information about a property, or detail on the decisions farmers should be making. But farmers have told us they want things simple to use, fast and easy to understand the result,” he said.

“The best tech fits easily into their farm business and supports their decisions.”

It’s all about the data

Since about 2010 there has been an explosion in the interest for data-driven agriculture to inform farm decisions and in automation to carry out those decisions.

You need a lot of accurate data to do good AI. Data from sensors, satellites and drones on farms open up myriad possibilities. Unfortunately, for now, data is the limiting factor in Australian agriculture. Australia has enough data sets on some aspects of agricultural productivity and not others. But data disaggregation is an obstacle: – meaning data is in different forms and is difficult to mesh together. Much of our data is also not publicly available.

But AI will soon be a force of innovation in agriculture in Australia. This will happen as more AI tools are developed to answer farming questions, as more data becomes available and aggregated, and as trust builds in these technologies.

As Dr Rahman says: “AI and machine learning can provide deep insights that suggest alternative practices with better results. That’s how AI can inspire innovation in agriculture.”

1 comments

  1. excellent, enjoyed reading this agro piece.

Leave a Reply