Background
An AgTech company wanted to develop a crop stress detection system using drone-captured multispectral imagery, but existing published detection models were trained on different crop varieties and growing conditions than those used by the company’s target customers.
The Challenge
It was unclear whether existing detection models could be adapted to the target crop and regional conditions, or whether the underlying spectral signatures of stress differed enough to require a new detection approach entirely.
Technological Uncertainty
It was not known in advance whether reliable stress detection was achievable for this crop variety and regional soil and climate conditions using available imagery, or what spectral indices and model architecture would be required.
Experimental Development
The team systematically tested multiple spectral index combinations and model architectures against field-validated ground-truth stress measurements collected across the growing season, measuring detection accuracy and false-positive rates for each configuration.
What Failed?
Standard spectral indices developed for other crop varieties showed poor correlation with ground-truth stress measurements in field trials, requiring the team to investigate custom spectral index combinations specific to this crop’s physiology.
Technological Advancement
The team developed a custom spectral index and detection model validated against field ground-truth data, generating new technical knowledge about stress detection for this crop variety under the target region’s growing conditions.
Potentially Relevant SR&ED Activities
- Systematic testing of spectral indices and model architectures against field data
- Field validation trials correlating imagery with ground-truth stress measurements
- Detection accuracy and false-positive benchmarking across model iterations
What Would Generally Not Qualify
Deploying the finalized, validated detection model for routine seasonal monitoring on similar fields would not itself qualify as further eligible development.
Documentation
Field trial data correlating imagery to ground-truth measurements, model accuracy benchmarks across iterations, and technical notes on index development would support this claim.
