Terres Inovia is an applied research institute in the agricultural sector, dedicated to supporting producers in increasing the diversification and sustainability of oilseed and protein crop cultivation. In order to effectively combat a pest, the flea beetle, in rapeseed crops, the institute identified an urgent need to automate its larval counting process, which was previously done manually: a long and laborious task.

To simplify and improve the efficiency of its protocols, the Terres Inovia institute wanted to automate the process of counting flea beetle larvae. The goal was not only to train a model to automate the process, but also to make it available to its teams for testing, with a view to subsequently opening it up to farmers and professionals in the sector. To do this, it was essential to leverage the existing dataset to train an artificial intelligence model capable of automatically counting larvae from images.

To meet this need, TRIMANE implemented several key steps, using various cutting-edge technologies and methods. The first step was to generate training images from the dataset provided by Terres Inovia to create a set of training images. The images were then adapted to match the model’s input requirements. Finally, the images were enriched to improve the robustness of the AI model.

The team deployed advanced tools, such as Azure Machine Learning for fine-tuning the YOLO v8 model, as well as a web application developed in Flask to enable technicians to use the trained model. The implementation of the project required the involvement of two data experts: a project manager and a consultant specializing in data science, enabling effective management and targeted technical expertise.

“We have continued to develop this tool, with new model training, as we had planned together from the outset. This year, we plan to make it available to external users for the fall season.

What I particularly appreciated about this project was the diversity of skills involved: machine learning project management, model training, development of a tool with a user interface, and implementation of an IT infrastructure enabling the continuous evolution of the model.” – Jean-Eudes HOLLEBECQ, New Agricultural Technologies Engineer D2IN at Terres Inovia.

The results obtained were significant: the model achieved a mean Average Precision (mAP) of 82%, with a difference of +/- 5% between the automated count and the actual count. This accuracy greatly improved the reliability of Terres Inovia’s tests. The switch to an automated solution should make the technicians’ work easier, but also paves the way for wider use of the system, including by professionals in the sector (technical specialists) and potentially farmers.

Terres Inovia’s internal satisfaction was evident, particularly with regard to the implementation of the Azure Machine Learning environment and the performance of the model. Terres Inovia appreciated the performance achieved and the creation of a functional web application for technicians. This project, carried out with TRIMANE, marked a significant step forward in the modernization of their processes.

“We benefited from the solid and varied skills of the TRIMANE team on this innovative project. Communication between our business teams and TRIMANE’s IT teams was very smooth, the project was very well managed and quickly resulted in an operational application. The TRIMANE teams understood our issues and needs very well, and our discussions enabled us to make progress together. This application is the successful result of a collaborative effort between our two organizations.” – Julie AUQUE, Data and Digital Innovation Department at Terres Inovia.