Crowdsourcing data and analysis competitions is a way to connect specialists such as crop physiologists with computer and data scientists, bringing new skill sets to meet grain industry challenges.

When Professor Scott Chapman begins a new academic year with crop physiology students, he always tells them they are privileged to have chosen a career in agricultural science.

“Agricultural science is one of the oldest science disciplines and we are tasked with applying the best new technologies to develop solutions to complex issues for growers,” Professor Chapman says.

However, only a limited number of students may choose this career path, especially as our society becomes increasingly urbanised and disconnected from the farmers who feed us.We need to innovate to cast our net wider.

That is exactly what an international team has done through a competition to entice data scientists into the agricultural space.

“Data competitions are a popular approach to crowdsource new data analysis methods for general and specialised data science problems,” Professor Chapman says.

“Competitions appeal to scientists researching in this area and can really drive innovation as the results are recognised achievements.”

The Global Wheat Challenge, managed by Dr Ian Stavness (University of Saskatchewan), was run in 2020 and 2021 to find more-robust solutions for wheat head detection using field images from different regions.

The Global Wheat Head Detection dataset used for the competition contained 6515 high-resolution RGB images representing 275,187 wheat heads from 16 institutions across five continents and 12 countries. The GRDC project directly supported labelling and competition activities in France (National Research Institute for Agriculture, Food and the Environment, INRAe and Arvalis) and Japan (the University of Tokyo).

Broadly speaking, the aim was to create an algorithm to detect individual wheat heads in field images to estimate head density – which is one of the main yield components of wheat and which underpins our understanding of genetic and agronomic responses.

The first competition attracted 2245 competitors and the second 432.

“We determined the winning challenge solutions in terms of their robustness when applied to new datasets,” Professor Chapman says.

The competition was supported by a global consortium of research institutes including GRDC through two projects: UOQ2002-008RTX ‘Machine learning applied to high-throughput feature extraction from imagery to map spatial variability’ and UOQ2003-011RTX INVITA ‘A technology and analytics platform for improving variety selection’.

Next iteration

Professor Chapman is the director of the University of Queensland node for the GRDC strategic partnership in Analytics for the Australian Grain Industry (AAGI), which is now collaborating with a larger global consortium to further automate the machine-learning-driven phenotyping of wheat.

“Semantic segmentation is a machine vision task that aims to classify each pixel as belonging to a type of object – for example, leaf, stem or head of a wheat plant,” he says.

Semantic segmentation methods have revolutionised the development of agricultural automation. However, the availability of labelled samples limits the training and evaluation of new semantic segmentation deep learning methods.

AAGI is supporting the labelling of more than 1000 new images from the consortium of universities (led by Dr Andreas Hund at ETH Zurich) to create a larger public resource for machine learning.

This AAGI project is contracting HiPhen, a company with expertise in high-throughput plant phenotyping, data acquisition and deep learning image analysis solutions applied to agriculture, to annotate the wheat image library.

“This will be a useful foundational resource to support future wheat research, trait identification, disease detection and much more.”

More information: Professor Scott Chapman, scott.chapman@uq.edu.au

Global Wheat dataset, global-wheat.com/gwfss.html

Read the journal paper