ABOUT THE PROJECT
AgriSpectra-Q
An industrial-oriented hyperspectral crop-intelligence and decision-support platform. Transforms real EnMAP Earth observation data into ranked inspection priorities — not a disease diagnostic, but an evidence-led decision-support signal.
CORE OPERATING PHILOSOPHY
Four-Step Workflow
DETECT
Real EnMAP L2A hyperspectral GeoTIFF data (224 bands, 30 m/px) is ingested with memory-aware windowed processing and NoData handling.
PRIORITISE
Spectral anomaly scores are computed, then scene-relative percentile thresholds generate a risk raster and a ranked priority map.
INSPECT
Connected high-priority zones are extracted, ranked, and converted to georeferenced zone cards with inspection recommendations.
VERIFY
Field teams receive ranked zone cards with spectral evidence. All findings require independent ground-truth verification.
RESULT MODES
Two Separate Result Modes
Live Matrix
A real-time, windowed, georeferenced spectral-anomaly analysis run directly on three EnMAP GeoTIFF scenes. Creates risk rasters, priority rasters, connected zones, GeoJSON, spectral evidence, and inspection-budget outputs.
Industrial Validation Study
A locked six-model comparison evaluated on 3 EnMAP scenes using spatially separated train/validation/test splits, 5 random seeds, frozen test predictions, and paired bootstrap analysis (n = 10,000 replicates).
ENMAP DATA
Scenes Processed
| Scene | Dimensions | Valid Pixels | NoData % | CRS | Proc. Time | HP Zones |
|---|---|---|---|---|---|---|
| Scene 01 (DT0000205230) | 1,153 × 1,198 | 1,028,176 | 25.56% | EPSG:32753 | 37.14 s | 407 |
| Scene 02 | 1,210 × 1,244 | 1,006,261 | 33.15% | EPSG:32645 | 111.41 s | 864 |
| Scene 03 | 1,152 × 1,214 | 1,047,911 | 25.07% | EPSG:32636 | 49.11 s | 438 |
Total live processing time: ~197.66 s · Run ID: AGRQ-LIVE-20260916-132530-587fc9
PRIORITY SYSTEM
Priority Zone Categories
Zone categories are scene-relative percentile thresholds for operational screening. They are not validated biological severity levels.
AI ARCHITECTURE
AgriSpectra-Q Model
RF-first residual architecture with grouped out-of-fold residual learning, compact spectral intelligence, Mahalanobis-oriented research components, and an adaptive residual gate. The nonlinear feature map uses quantum-inspired computational logic within a hybrid quantum-classical research layer.
Quantum Component Clarity
Correct framing: "Quantum-inspired feature transformation within a hybrid quantum-classical research layer." No quantum hardware result, no quantum speedup, no demonstrated quantum advantage.
Statistical Position
AgriSpectra-Q achieves the highest numerical mean F1 (0.963985) among all six evaluated systems. However, its advantage over HSI-RF is only +0.0008447 with a 95% CI of [–0.0012, +0.0027], which crosses zero. Statistical superiority is therefore not established.
FUTURE DEVELOPMENT
Roadmap
Phase 1: Raw-Input Validation
Preserve raw spectra, coordinates, and wavelength metadata. Re-run from raw input without test correctness.
Phase 2: External Geographic Validation
Evaluate a blind fourth EnMAP scene frozen from the development process.
Phase 3: Field Validation
Collect field polygons and agronomist labels to replace spectral-proxy with biological truth.
Phase 4: Temporal Intelligence
Multi-date analysis for persistence, trend, and temporal anomaly velocity.
Phase 5: Operational Pilot
Measure inspection recall, travel cost, false-alarm burden, and relative decision utility.
Phase 6: Production Deployment
Cloud-scale storage, async jobs, auth, monitoring, drift detection, and versioning.
TARGET AUDIENCE
Intended Users
Scientific Boundaries
AgriSpectra-Q identifies spectral-anomaly priority candidates for field inspection. It does not diagnose disease, pests, or biological stress. It does not claim field validation, measured financial ROI, statistically significant superiority over HSI-RF, or quantum advantage. All priority zones require independent field verification.