Built to beat noise in the real world.
Scatter correction, moisture orthogonalization, denoising, and sub-pixel detection—anchored to lab truth.
Preprocessing
Robust signal preparation for reliable detection in field conditions.
Scatter Correction
Real-world spectra are affected by physical scattering from particle size, surface texture, and sample geometry. Our preprocessing pipeline addresses these challenges:
- SNV (Standard Normal Variate) — Removes baseline shift and path-length variation
- MSC (Multiplicative Scatter Correction) — Corrects multiplicative scatter effects
- EMSC (Extended MSC) — Advanced correction separating scatter from chemical absorbance
Moisture Handling
Water bands dominate NIR spectra in undried hay and green pasture. We use orthogonalization methods to remove moisture interference while preserving target signals:
- OSC (Orthogonal Signal Correction) — Removes variance uncorrelated with targets
- ROSC (Robust OSC) — Enhanced stability across moisture ranges
- Calibration sets spanning real-world moisture levels ensure robust performance
Denoising & Features
Extracting meaningful signal from noisy field data.
Hyperspectral Cube Processing
For drone-collected imagery, we apply dimensionality reduction that separates signal from noise:
- MNF (Minimum Noise Fraction) — Standard noise-ordered transform
- OKMNF — Optimized kernel variant for improved separation
- Preserves subtle spectral features while removing sensor and atmospheric noise
Spectral Denoising
For point spectra from handheld analyzers:
- Wavelet denoising — Multi-scale noise removal preserving sharp spectral features
- Adaptive thresholding based on noise characteristics
- Preserves diagnostic absorption bands
Spectral-Spatial Fusion
Combining spectral signatures with spatial context improves detection in heterogeneous scenes:
- GLCM texture — Gray-level co-occurrence matrices capture patch structure
- Spatial context helps distinguish isolated plants from patches
- Improves accuracy in mixed vegetation scenarios
Detection & Calibration
Lab-grounded models tuned for real-world performance.
Sub-Pixel Detectors
Finding targets that don't fill an entire pixel is the key challenge for early detection. Our detection algorithms are designed for this scenario:
- CEM (Constrained Energy Minimization) — Optimal filter for known target spectra
- ACE (Adaptive Coherence Estimator) — Background-adaptive detection
- Matched Filter — Classic approach with scene adaptation
- All methods tuned with clean lab spectra and adapted to each scene
Calibration Approach
Robust detection requires rigorous calibration with known samples:
- Spiked sample sets — Known "blades/kg" concentrations across moisture levels
- Lab reference spectra — Pure target spectra under controlled conditions
- Field validation — Ground-truth verification in real conditions
- Focus on low-cover, mixed-pixel signals (e.g., single blades in bales)
Want the full technical details?
Download our technical brief for a deeper dive into methods, validation results, and citations.
Download the technical brief (PDF)