Research & Evidence
Scientific protocol, validation evidence, and current limitations of the AgriSpectra-Q industrial PoC. All results are frozen — no live run affects these figures.
DATASET & EVALUATION EVIDENCE
What Evidence Exists
BENCHMARK METRICS
Key Numerical Results
VALIDATION PROTOCOL
Eight-Step Scientific Protocol
Scene preparation
Three real EnMAP L2A GeoTIFF scenes. Windowed streaming with constant memory footprint. NoData masking.
Spectral anomaly proxy
RMS standardised deviation over 32 evenly-spaced bands. Scene-relative percentile thresholds (P50/P80/P95). Unsupervised — no disease labels.
Spatial evaluation splits
Spatially separated train/validation/test split per scene. No pixel-level leakage between splits.
Five random seeds
Seeds 11, 22, 33, 44, 55 per model per scene. 90 total evaluation runs. Mean and std computed per metric.
Frozen test predictions
Test-set predictions locked before analysis. No post-hoc adjustment of thresholds or model parameters.
Paired bootstrap
n = 10,000 replicates. 95% confidence interval for pairwise difference vs HSI-RF. CI reported verbatim.
Calibration
Post-hoc Platt scaling. Expected Calibration Error (ECE) reported before and after calibration.
Ablation study
Quantum-inspired component isolated. Measurable contribution confirmed on Scene 01. Not a quantum hardware claim.
CURRENT LIMITATIONS
What This PoC Does Not Provide
No field validation — no field-labelled disease or pest target exists.
No blind fourth EnMAP scene — all scenes were used during development.
No complete six-model LOSO evaluation — leave-one-scene-out not yet verified.
Benchmark target is a Spectral Anomaly Proxy — not an independently labelled biological outcome.
No proven quantum advantage — the quantum-inspired layer provides measurable contribution but no quantum hardware was used.
No measured financial ROI — inspection cost savings are not verified against operational baselines.
Wavelength metadata unavailable in GeoTIFF — band indices reported, physical wavelengths not verified.
Priority thresholds are scene-relative percentiles — not absolute disease severity levels.
Processing time varies per scene — not suitable for real-time operational use without further optimisation.
The current PoC has no field validation, no field-labelled disease or pest target, no blind fourth scene, no complete six-model LOSO evaluation, no proven quantum advantage, and no measured financial ROI. All priority zones require independent field verification.