Machine learning platform for detecting belowground crop pests using root volatile patterns

Overview

This solution applies machine learning to root volatile emissions to detect and predict belowground herbivory in crops. The approach is analogous to "smelling" a barcode: plants attacked by root-feeding pests release specific volatile compounds in distinctive concentrations and ratios that differ from uninfested plants. By training classifiers on these chemical signatures, the system can determine whether pests such as western corn rootworm larvae are present in the soil.

The technology addresses a critical gap in integrated pest management. Belowground pests are difficult to detect until significant root damage has occurred, leading to yield losses before interventions can be applied. Early, non-destructive detection through volatile analysis offers a pathway to more timely and targeted pest management decisions, reducing unnecessary pesticide applications and supporting precision agriculture workflows.

Technical specifications

Detection approach:

• Custom-designed belowground volatile collection probe captures root emissions directly from the rhizosphere
• Gas chromatography/mass spectrometry (GC/MS) separates and identifies individual volatile compounds
• Chromatographic peaks are matched against an in-house compound library to build structured datasets
• Open-source chromatography software processes raw data for downstream analysis
• Machine learning algorithms developed in Python classify samples as infested or non-infested based on volatile profiles

Validated performance:

In a completed study, root volatile data from 56 infested and 71 control maize plants was used to train and test a support vector machine classifier. Cross-validation demonstrated 78% accuracy in predicting the presence or absence of western corn rootworm larvae. Ongoing work aims to increase this accuracy by expanding sample size and testing additional algorithms.

Key advantages:

• Non-destructive sampling that does not require uprooting plants
• Species-level detection potential based on unique volatile signatures
• Compatible with existing analytical chemistry infrastructure

Technology readiness level

The technology is at an early-to-mid stage of development. Proof-of-concept validation has been completed with a 78% classification accuracy using a support vector machine approach on a limited sample set. Future validation will involve collecting hundreds of root volatile samples from maize under controlled infestation conditions over a two-year period, followed by systematic evaluation of multiple machine learning algorithms to optimize predictive performance. Field-deployable sensor integration remains a longer-term goal, as current GC/MS-based analysis is not yet practical for in-field use. The research team is seeking partners to support algorithm development and funding for key personnel to advance the technology toward commercial readiness.