AI in agriculture research

Romain Royer

,

Marketing Cooperator

February 13, 2024

5 min

Artificial intelligence is on the verge of true decision-making assistance in agriculture research. Concocting plant breeding software for 40 years, we are weighing our words here.

Far from OpenAI and ChatGTP, let’s dive into the deep world of machine learning. You will discover the potential of an algorithm that integrates the dataset of a whole R&D department for the upmost mission: Help plant breeders take decisions.

“Well” or “less” adapted to the breeding objectives

In his research work, Rony Charles has trained a neural network on more than 20 thousand microplots of hybrid Corn. Now choose a new plot among your Multi-Environment Trials for which you know the soil, weather indicators, and field observations. Give all this information to the trained model and the latter will classify that plot as “well adapted” or “less adapted” to the conditions in which the plot has grown.

Now you wonder: How accurate is the result?

It guesses an expert’s analysis with 95% accuracy, which means 19 times on 20!

🟢🟢🟢🟢🟢| 🟢🟢🟢🟢🟢 | 🟢🟢🟢🟢🟢| 🟢🟢🟢🟢🔴

I told you it’s mo-men-tous.

Rony is the best person to explain this poster entitled “Genotype Classification Using an Artificial Neural Network on Multivariate Phenotyping Datasets”. Hear what he says about the poster presented at the last Plant and Animal Genome Conference in January 2023:

Interview of Rony Charles about Artificial Network on phenotyping datasets

Artificial Intelligence and plant breeding and software

Rony Charles is an AI enthusiast. He dreams of developing a game-changing model for agriculture. This is why, after his MSc, he chose Université Côte d'Azur and its renowned artificial intelligence laboratory: 3IA. He is now preparing a PhD thesis under the supervision of Prof. Andrea Tettamanzi, expert in evolutionary algorithms,fuzzy logic, and soft computing (or computational intelligence).  

RAGT is at the forefront of agricultural research innovation. They deliver farmers worldwide with ones of the best varieties of Forages, Corn and Cereals. Bruno Claustres leads the data & digital innovation of the agriculture research entity RAGT 2n. RAGT provided Rony with the necessary data to train and test the AI models and guided him to elaborate a genuinely useful artificial intelligence tool.

Doriane develops software for agronomists in various organizations around the world. Their technology supports plant R&D departments to centralize, secure and valorize their databases. Breeders,testing managers and variety product developers rely on Doriane’s experience to monitor their plant breeding KPIs (Key Performance Indicators). They appreciate using their state-of-the-art tools to avoid mistakes and speed-up their research.

Needless to say, the pooling of all these skills and resources provides a unique breeding ground to develop a model that supports agricultural researchers!

The Artificial Neural Network analyzes agriculture research data

Helping plant breeders achieve their goals

What is the point of developing an AI model to test hybrids? What technology could really support agricultural research? Rony asked himself these questions for some time before realizing what matters most: the breeding objectives.

The aim of a plant breeder is to find a variety that performs well in a specific market. For instance, RAGT is looking for an early, drought resistant maize ensuring high yields. These characteristics correspond to the needs of farmers in certain regions of the Hungarian market.

To train the AI model, Rony has used the analysis of RAGT experts to label all the data. Every plot has been labeled upon the agronomist evaluation: “well adapted” or “less adapted” to the growing conditions, according to the experimentation goals.

Now that the models are trained, what happens when you provide a new plot with all the genotype, phenotype, soil and weather data? The model can guess if the genotype of that plot is adapted to the conditions in this field, according to the breeding objectives!

trait notation agriculture corn artificial intelligence
Taking notations in a corn field with RnDExp™ Mobile

Training the AI on a multi-variate, multi-location phenotyping dataset

Artificial intelligence accuracy depends mainly on the quality of available data and the appropriate algorithm. For this study RAGT 2n provided 5 years of several thousands of hybrids tested in multiple locations.

The data is organized in several hundreds of Multi-Environment Trials (METs). Every MET represents the experimentation of a research team from sowing to harvest. RAGT researchers are using RnDExperience™ plant breeding software.

3 kinds of data have been used to train the model:

  • Phenotype: Traits describing the hybrid maize. We had notably: Precocity,initial vigor, plant height, length of the ears, harvested yield and moisture,and disease occurrence.
  • Soil information, where the plant has grown:location in the field, soil characteristics such as total available water,clay, fine and coarse sand, etc.
  • Weather information such as rainfall, solar radiation, temperature, wind etc.

So, the data set is a mixture of those three types of data, but we have tried several scenarios: How accurate is the model if we provide only phenotype data? Is it better with soil or weather?

In the end, all models have given better results with phenotype + soil + weather data. It seems obvious but that confirms the pertinence of the whole study’s methodology.

Artificial intelligence model accuracy AI neural network
Accuracy of 7 AI models on 4 different datasets

95% accuracy with an Artificial Neural Network

After experimenting with various models, it has been decided to keep 7 of them. Those models have been trained on 70% of the dataset, and then tested on the 20% remaining. The best accuracy has been obtained with an AI algorithm called “Artificial Neural Network”.

But don’t forget Rony’s wise sentence: “There’s no perfect model!”

In this study, 95% is a pretty good accuracy: Over 100 examples, on 95 instances both the machine and the expert come to the same conclusion. Which actually means that the expert can rely on the machine 95% of the time to decide if their material in that specific environment is well adapted based on the breeding program goals.

Rony explains also that training an artificial neural network takes more time than the other algorithms. So, if you have limited resources or if you have less data, a logical regression may be useful too.

RAGT agriculture reearch field
Harvesting a plot of Corn in an RAGT testing field ©RAGT les-semeurs-ragt.fr

The future of plant breeding software

The truth is that this AI model is not directly usable by agronomists: Training the data has required several weeks of work (notably the labelling process) and the algorithm takes hours to evaluate the plots. But this may be the closest we have ever come to creating an AI tool that helps agricultural researchers in their daily work.

Everyday plant breeders have questions about some plots. They are in doubt, not knowing whether the crop is adapted or not. Soon they will be able to ask an artificial intelligence to help them make such decisions.

The collaboration goes on between RAGT, Doriane, UCA. Hopefully it will bring about more tools to help researchers create varieties that match with climate-smart plant breeding objectives!

Did you find this article useful? Share it!

twitter iconlinkedin iconfacebook icon

Discover our related webinar

Do not miss this webinar

Related content

Article

New genomic techniques approved in EU

See more

Article

3 Steps to standardize processes in agriculture research

See more

Go futher with this Ebook