Earth Observation · isp-uv-es
rtm_emulation
A benchmark dataset for training fast machine learning emulators to replace atmospheric radiative transfer models.
What rtm_emulation is
rtm_emulation is an atmospheric radiative transfer model (RTM) emulation benchmark developed by the Image and Signal Processing group at Universitat de València (isp-uv-es). Atmospheric RTMs are foundational tools in remote sensing, numerical weather forecasting, and climate science, but their heavy computational costs prevent direct use in operational pipelines. Standard look-up table (LUT) interpolation approaches require massive storage, struggle to scale with hyperspectral instrumentation, and lack flexibility across changing atmospheric conditions.
This dataset provides training simulations and evaluation benchmarks to train statistical regression algorithms and neural surrogates that replicate complex deterministic RTM behaviour. By replacing computationally intensive forward simulations with fast surrogates, it aims to facilitate real-time satellite data processing, physics-based inversion routines, and the atmospheric correction of hyperspectral imagery.
How it was made
The dataset is curated using Latin Hypercube Sampling to cover an approximately 10-dimensional input variable parameter space while avoiding ill-posed configurations. Forward simulations produce hyperspectral output functions across several thousand spectral intervals. Data are split across two core experimental scenarios: atmospheric correction (Scenario A) and carbon dioxide column retrieval (Scenario B).
Each scenario contains varying training subsets (such as 2,000 samples) provided in HDF5 format, storing atmospheric transfer functions (LUTdata), input parameter matrices (LUTHeader), and wavelength grids (wvl). Test datasets are supplied as tabular CSV files split into interpolation (Track 1) and extrapolation (Track 2) evaluation benchmarks.
Key facts
- Domain
- Earth observation, climate physics, remote sensing
- Primary data formats
- HDF5 (.h5), CSV (.csv)
- Input dimensionality
- Approximately 10 parameters
- Output dimensionality
- Several thousand spectral intervals
- Scenario A spectral range
- 400 to 2500 nm at 5 cm⁻¹ resolution
- Scenario B spectral range
- 2000 to 2100 nm at 0.1 cm⁻¹ resolution
- Evaluation metrics
- Spectrally-averaged mean relative error (MREλ) and runtime
What you can use it for
- Hyperspectral Atmospheric Correction
- Train emulators to predict path radiance, direct and diffuse solar irradiance, and atmospheric transmittance across 400–2500 nm to retrieve surface reflectance from satellite radiance.
- Carbon Dioxide Column Retrieval
- Model interactions across atmospheric gas layers to accurately predict top-of-atmosphere radiance in the 2000–2100 nm CO2 absorption window at high resolution.
- Surrogate Radiative Transfer Modelling
- Replace memory-heavy look-up table interpolation with lightweight machine learning regression models for fast physics-based remote sensing inversions.
- Generalisation and Extrapolation Benchmarking
- Evaluate how machine learning models perform when predicting radiative transfer variables under unseen or out-of-distribution atmospheric and geometric conditions.
Explore the data
No queryable preview
This dataset is not available in a queryable form — usually an unusual file format or a repository too large to convert. The original source may still be readable.Strengths and limits
Good at
- Uses Latin Hypercube Sampling to ensure robust coverage across input parameter spaces.
- Provides dedicated interpolation and extrapolation reference sets to evaluate generalisation limits.
- Includes multi-resolution spectral configurations spanning broad-spectrum and narrow-band absorption features.
- Supplies ready-to-load examples for common scientific environments, including Python, MATLAB, and R.
Watch out for
- Extrapolation scenarios (tracks A2 and B2) exhibit significantly higher relative errors across evaluated models than interpolation tracks.
- The high input and output dimensionality presents substantial model training complexity.
- Performance in atmospheric correction is sensitive to deep water vapour (H2O) absorption bands, which are masked during metric evaluation.
Questions
Can I look inside rtm_emulation without downloading it?
Not here: it is not available in a queryable form. Its description and statistics are still shown.
What licence is rtm_emulation under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What is the difference between Scenario A and Scenario B?
Scenario A focuses on atmospheric correction across the 400–2500 nm range at 5 cm⁻¹ resolution, predicting transfer functions such as path radiance and irradiance. Scenario B focuses on CO2 column retrieval across the 2000–2100 nm absorption band at 0.1 cm⁻¹ resolution, targeting top-of-atmosphere radiance.
How are interpolation and extrapolation tested?
Each scenario contains two tracks: Track 1 (refInterp) tests standard interpolation performance within the sampled domain, whereas Track 2 (refExtrap) tests extrapolation capabilities under out-of-distribution parameters.
What metrics are used to benchmark model accuracy?
Models are evaluated using the spectrally-averaged mean relative error (MREλ). For Scenario A, error is calculated between retrieved and reference reflectance excluding deep H2O absorption intervals; for Scenario B, error is computed directly on top-of-atmosphere spectral radiance.
Is runtime factored into the challenge ranking?
No, runtime is reported in seconds to assess computational efficiency, but the final competition ranking is calculated entirely using a weighted average of individual track accuracy ranks.
Written from rtm_emulation's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.