Ai-powered detection model for new zealand's most unwanted invasive insect pests

Overview

This project develops a computer vision and deep learning model capable of detecting eight of New Zealand's most threatening invasive insect pest species from everyday mobile phone images. The target species include Brown marmorated stink bug, Bagrada bug, Yellow-spotted stink bug, Fall armyworm, Gypsy moth, Oriental fruit fly, Queensland fruit fly, and Mediterranean fruit fly. Early and accurate identification of these pests is critical for biosecurity readiness, as their arrival could cause multi-billion dollar impacts on agriculture, native ecosystems, and New Zealand's premium residue-free production systems.

The solution is designed for practical field use, enabling farmers, biosecurity officers, and the public to capture images with smartphones and receive rapid species identification. By combining public and private datasets, the model aims to achieve at least 90% average accuracy across all eight species at various life stages.

Technical specifications

Core approach:

• Deep learning object detection model trained to classify eight target pest species against each other and against thousands of images of other insects, including visually similar native species
• Training dataset combines Plant & Food Research data, collaborator contributions, and relevant public datasets, targeting at least 200 examples per pest species
• Built using state-of-the-art object detection methods on an internal machine learning platform
• Designed to handle variable image resolutions, lighting conditions, natural backgrounds, and amateur photography quality

Prior validation results:

• A previous 2-class model achieved 96.5% mean average precision for Brown marmorated stink bug detection and 93.5% for other insects across 9,327 images
• Training data included 2,436 BMSB instances across all life stages and 3,774 instances of visually similar or related insects

Target performance:

• At least 90% average accuracy across all eight pest species
• Robustness across life stages and field conditions

Technology readiness level

The project builds on a previously validated 2-class detection model with demonstrated accuracy above 93% on real-world image datasets. The current proposal extends this proven approach to eight species using a larger, more diverse dataset. The next stage involves dataset assembly, model training, and validation against the 90% accuracy target. The technology is at an early-to-mid development stage, moving from proof-of-concept toward a field-ready biosecurity detection tool.