SVH-BD : Synthetic Vegetation Hyperspectral Benchmark Dataset for Emulation of Remote Sensing Images

Abstract: This dataset provides a large collection of 10,915 synthetic hyperspectral image cubes paired with pixel-level vegetation trait maps, designed to support research in radiative transfer emulation, vegetation trait retrieval, and uncertainty quantification. Each hyperspectral cube contains 211 bands spanning 400--2500 nm at 10 nm resolution and a fixed spatial layout of 64 \times 64 pixels, offering continuous simulated surface reflectance spectra suitable for emulator development and machine-learning tasks requiring high spectral detail. Vegetation traits were derived by inverting Sentinel-2 Level-2A surface reflectance using a PROSAIL-based lookup-table approach, followed by forward PROSAIL simulations to generate hyperspectral reflectance under physically consistent canopy and illumination conditions. The dataset covers four ecologically diverse regions -- East Africa, Northern France, Eastern India, and Southern Spain -- and includes 5th and 95th percentile uncertainty maps as well as Sentinel-2 scene classification layers. This resource enables benchmarking of inversion methods, development of fast radiative transfer emulators, and studies of spectral--biophysical relationships under controlled yet realistic environmental variability.

• The paper introduces SVH-BD, a benchmark dataset of 10,915 synthetic hyperspectral image cubes with dense pixel-level vegetation trait annotations.
• It employs a robust methodology combining Sentinel-2 preprocessing, PROSAIL-based spectral inversion, and forward radiative transfer simulation to generate realistic spectra and uncertainty maps.
• The resource enables rigorous evaluation of RTM emulators, vegetation trait retrieval algorithms, and uncertainty-aware models for remote sensing applications.

SVH-BD: A Synthetic Benchmark Dataset for Vegetation Hyperspectral Image Emulation

Introduction

The "SVH-BD: Synthetic Vegetation Hyperspectral Benchmark Dataset for Emulation of Remote Sensing Images" (2603.28390) introduces a comprehensive collection of 10,915 synthetic hyperspectral image (HSI) cubes, each paired with dense, pixel-level annotations of vegetation biophysical traits. Generated through a pipeline coupling satellite data preprocessing, radiative transfer model (RTM) inversion, and forward hyperspectral simulation, SVH-BD aims to directly address the persistent lack of training and benchmarking datasets for remote sensing inversion methods and RTM emulator development. The resource is specifically tailored for applications involving radiative transfer emulation, vegetation trait retrieval, and uncertainty quantification, providing physically realistic, annotated hyperspectral data where such ground truth is typically unavailable in real satellite archives.

Data Generation Methodology

SVH-BD employs a multi-step synthetic data pipeline rooted in established physical modeling frameworks. The dataset construction comprises:

1. Sentinel-2 Acquisition and Preprocessing: Level-2A products from Sentinel-2 MSI are sourced for four ecologically distinct regions (East Africa, Northern France, Eastern India, Southern Spain), and harmonized at 20 m GSD. Multispectral data mosaics are spatially cropped to 64×64 pixels to standardize all input scenes.
2. Spectral Inversion: Using a PROSAIL-based LUT, multispectral reflectance is inverted at each pixel to estimate vegetation biophysical parameters. Latin Hypercube Sampling with tight physiological constraints ensures biologically plausible parameters. The inversion process admits the ten lowest-RMSE LUT entries for every pixel, providing both parameter estimates (ensemble median) and empirical uncertainty quantiles (5th and 95th percentiles).
3. Forward RTM Simulation: Retrieved parameters are propagated through PROSAIL, integrating PROSPECT-D for leaf optical property simulation and SAIL for canopy-level radiative transfer. This enables generation of continuous, physically consistent reflectance spectra across 211 bands (400-2500 nm, 10 nm spacing) per sample.
4. Scene Contextualization: Additional Sentinel-2 scene classification layers clip each hyperspectral cube, supporting categorical landscape analysis.

Each region is associated with a single representative soil spectrum sourced from the ICRAF-ISRIC VNIR spectral library, ensuring internally consistent but somewhat idealized soil-vegetation reflectance simulation.

Dataset Characteristics

Each SVH-BD sample comprises:

• A $64 \times 64 \times 211$ hyperspectral reflectance cube.
• A 16-channel trait map detailing leaf, canopy, and observation parameters, including Cab, Car, LAI, LIDFa, soil type, and solar/view geometry.
• 5th and 95th percentile uncertainty maps quantifying inversion ambiguity.
• A categorical Sentinel-2 scene classification map with accompanying metadata.

The dataset encompasses over 10,000 uniquely parameterized image cubes, ensuring a high degree of spectral, biochemical, and structural variability. Regional diversity is explicitly modeled, with each geographical domain reflecting domain-specific parameter distributions and realistic canopy-soil optical coupling.

Numerical Strength and Claims

A notable strength of SVH-BD lies in the explicit provision of both pixel-level parameter estimates and non-parametric uncertainty bounds. Unlike prior simulated datasets or raw satellite data, SVH-BD supplies biophysical annotations (e.g., chlorophyll content $0-160\, \mu g/cm^{2}$, LAI $0-10$), which are otherwise infeasible to acquire at scale from orbital sensors. The LUT inversion approach, supported by a library of 50,000 PROSAIL spectra, is robust to high-dimensional parameter degeneracies and enables quantification of retrieval ambiguity under real spectral noise and spatial heterogeneity.

Another key claim of the resource is its physical and physiological realism: parameterizations are physiologically constrained (e.g., Cab-LAI coupling) and filtered for spectral plausibility (e.g., exclusion of green peaks below 547 nm). This stands in contrast to naïve sampling of parameter space, which often generates non-vegetative spectra.

The dataset’s spectral coverage (211 bands at 10 nm resolution) enables the emulation of spaceborne HSI imagers and supports machine learning tasks where detailed spectral information is required but rarely available with annotation.

Practical and Theoretical Implications

SVH-BD directly enables several research directions:

• RTM Emulator Training: The dataset is optimized for the development of ML-based forward/inverse RTM emulators, supporting architectures that require high-resolution spectral cubes and pixel-aligned trait labels.
• Trait Retrieval Benchmarking: Pixel-level trait ground-truth supports quantitatively rigorous evaluation and inter-comparison of existing and novel inversion algorithms.
• Uncertainty-aware Approaches: Provision of uncertainty maps enables the development of robust, uncertainty-quantifying retrieval and downstream models, a core requirement for operational decision-making in environmental monitoring.
• Land Cover-aware Analysis: Integration of land-cover classification enables analysis that jointly leverages spectral, biophysical, and categorical context.

On a theoretical level, systematic generation and benchmarking of emulators using this dataset may drive advances in the expressivity, calibration, and interpretability of ML models for RTMs, potentially informing the next generation of operational trait retrieval algorithms and their integration with upcoming HSI missions.

Limitations and Future Perspectives

The dataset has several recognized limitations. The ecological and bio-optical diversity is restricted to four regions and one soil spectrum per region, thus underrepresenting certain environmental gradients and edaphic effects. Soil-vegetation spectral decoupling, important for some retrieval diagnostics, cannot be systematically evaluated. The pipeline’s reliance on the forward-inverse PROSAIL chain introduces the known limitations of 1D RTMs (e.g., absence of explicit gap fraction or clumping effects) and may not perfectly reproduce the intricacies of real mixed pixels or heterogeneous canopies.

Nonetheless, SVH-BD establishes a scalable workflow that could be extended to broader spatial domains, additional soil libraries, and other canopy architectures. Incorporation of more advanced RTMs, functional trait perturbations, and higher-order uncertainty quantification methods (e.g., Bayesian RTM inversion) are natural avenues for expansion. Advancements in high-fidelity generative modeling (e.g., NeRFs, diffusion models for HSI) could be conditioned or benchmarked against this reference dataset in the future.

Conclusion

SVH-BD (2603.28390) represents a substantive step forward in the simulation-based benchmarking of vegetation hyperspectral imaging, offering detailed, physically plausible spectral cubes with explicit trait and uncertainty annotation. The dataset’s structure and documentation directly support ML research on RTM emulation, trait retrieval, and uncertainty quantification, while remaining transparent about key physical and algorithmic assumptions. Its design enables the rigorous, comparative evaluation of ML and RTM-based methods, and establishes a robust foundation for future methodological and dataset extensions in the remote sensing community.