FROZEN SCIENTIFIC BENCHMARK90 Runs · 6 Models · 3 Scenes

Benchmark Results

Six models evaluated across 3 real EnMAP scenes with 5 random seeds each. These are frozen scientific results — not live analysis outputs. For a live run, go to Intelligence → Run Analysis.

SCIENTIFIC INTERPRETATION

Competitive numerical result

AgriSpectra-Q has the highest mean F1 (96.40%) among all six evaluated systems.

Statistical superiority over HSI-RF not established

Bootstrap 95% CI [−0.0012, +0.0027] crosses zero. Not statistically significant.

Benchmark target is a Spectral Anomaly Proxy — not independently labelled disease or pest outcome.

96.40%

Mean F1

3 scenes × 5 seeds

99.47%

PR-AUC

Precision-Recall

99.87%

ROC-AUC

Discrimination

0.73%

Calibrated ECE

Prob. calibration

TABLE 1

Six-Model Benchmark

Mean over 5 seeds × 3 scenes

Model	Mean F1	PR-AUC	ROC-AUC
AgriSpectra-QOur System	96.40%	99.47%	99.87%
Adaptive Classical	96.34%	99.47%	99.86%
HSI-RF	96.31%	99.49%	99.87%
48-band XGBoost	95.22%	99.23%	99.80%
Spectral XGBoost	94.78%	99.24%	99.81%
Current Hybrid	89.98%	96.16%	98.75%

TABLE 2

AgriSpectra-Q — Per-Scene Results

Scene	F1	Precision	Recall	ROC-AUC
Scene 01 (DT0000205230)	98.47%	98.23%	98.73%	99.98%
Scene 02	95.42%	95.80%	95.10%	99.83%
Scene 03	95.30%	94.81%	95.83%	99.79%

KEY FINDINGS

What the Benchmark Tells Us

AgriSpectra-Q achieves 96.40% mean F1 — highest numerical result among all six evaluated systems.

Scene 01 (DT0000205230) delivers 98.47% F1, reflecting high spectral contrast in the Al Ain oasis region.

Calibration ECE of 0.73% after post-hoc scaling — well-calibrated probabilities for decision support.

At 10% inspection budget, the system achieves ~49% positive recall — ~5× better than random sampling.

Ablation studies confirm the quantum-inspired component provides measurable contribution.

Paired bootstrap CI [−0.0012, +0.0027] vs HSI-RF at n=10,000 replicates.

VALIDATION METHODOLOGY

5 Random Seeds

11, 22, 33, 44, 55

Statistical stability across initialisation variance

3 Real EnMAP Scenes

UAE & Gulf region

No simulated data — all results on actual EO imagery

Paired Bootstrap

n = 10,000 replicates

Rigorous significance testing for model comparisons

Run a live spectral-priority analysis

Apply the live engine to real EnMAP scenes and generate georeferenced priority zones.

Run Live Analysis