Published November 9, 2026 | Version 1.3.0

OpenTIR-PINN Benchmark v1.3.0: Multi-Atmospheric Radiative Transfer and Double-Quarantine Golden Target Dataset for Ill-Posed Temperature and Emissivity Separation

  • 1. ROR icon Universidade Federal do Rio Grande do Sul
  • 2. ROR icon University of Chile
  • 3. ROR icon Van Lang University

Description

OpenTIR-PINN Benchmark (Version 1.3.0)
Comprehensive open-science benchmark dataset and reproducible data pipeline for Physics-Informed Neural Networks (PINN), radiative transfer inversion, and operational Temperature and Emissivity Separation (TES) algorithms across multispectral (ASTER, 5 bands) and hyperspectral (HyTES, 256 bands) thermal infrared sensors.

Key Features & Upgrades in Version 1.3.0:
1. Complete Radiative Transfer Coupling & 330 Archived MODTRAN 5.2 Decks: All simulated radiances strictly satisfy the full surface-leaving boundary condition L_surf = epsilon * B(T_s) + (1 - epsilon) * L_down and Top-of-Atmosphere coupling L_TOA = tau * L_surf + L_up across 30 distinct MODTRAN 5.2 profiles (6 standard atmospheric models x 5 water vapor scaling factors, PWV from 0.21 to 6.77 g/cm²), shipped alongside all 330 executed MODTRAN input decks.
2. Operational Atmospheric-Correction Error Propagation (N4) & Sensor Noise: Network inputs are derived through operational atmospheric correction using estimated atmospheres perturbed by realistic precipitable water vapor (PWV) retrieval errors and sensor-specific NEΔT detector noise (0.3 K for ASTER, 0.2 K for HyTES).
3. Group-Disjoint 5-Fold Cross-Validation & Double-Quarantine Suite: Provides 231 curated development spectra grouped by physical specimen and spectral similarity (162 groups across 5 folds, N=5,460 training and N=2,156 out-of-fold validation samples in Fold 0), plus 7 immutable Golden Targets (urban concrete, urban asphalt, calcite, kaolinite, quartz, anorthite, and Abies concolor bark) evaluated across In-Distribution Atmospheres (Quadrant II, N=308) and Quarantined Winter Atmospheres (Quadrant IV, N=770).
4. Single Source of Truth & 32-Test Automated Validation Protocol: Includes a complete deterministic rebuild pipeline (pipeline/run_pipeline.py), cryptographic manifest (data/dataset_facts.json), 32-test integrity validator, and standalone 7-gate verification script (quickstart_demo.py).

Files

DATA_DICTIONARY.md

Files (72.5 MB)

Name Size
md5:f1d2dd1f3912a6e357afcc15d8360fb8
13.1 kB Preview Download
md5:f3f46bcdd5f50ada8bd669255f701fcc
61.7 kB Preview Download
md5:1ef5b544e2cca5a7dcd5801d29670262
33.3 kB Preview Download
md5:378a1d3cc954aef6f60fda5a531c0bd9
72.4 MB Preview Download

Additional details

Related works

Is supplement to
Dataset: 10.5281/zenodo.23571365 (DOI)
Preprint: 10.31223/X5MJ8J (DOI)
Is supplemented by
Software: https://github.com/atilioebg/OpenTIR-PINN (URL)

Software

Repository URL
https://github.com/atilioebg/OpenTIR-PINN
Programming language
Python
Development Status
Active