Verrelst et al. (2025) Surrogate Modeling in Optical Remote Sensing: A Review of Emulation for Vegetation and Atmosphere Applications

Identification

• Journal: Preprints.org
• Year: 2025
• Date: 2025-09-11
• Authors: Jochem Verrelst, Miguel Morata, José Luis García-Soria, Yilin Sun, Jianbo Qi, Juan Pablo Rivera
• DOI: 10.20944/preprints202509.0938.v1

Research Groups

Jochem Verrelst, Miguel Morata, Jose Luis Garcia, Yilin Sun, Jianbo Qi, Juan Pablo Rivera-Caicedo

Short Summary

This review comprehensively surveys recent developments in surrogate modeling (emulation) for vegetation and atmospheric radiative transfer models (RTMs) in optical remote sensing, highlighting methodologies, applications, and future challenges to address the computational cost of RTMs.

Objective

• To comprehensively survey recent developments in emulating vegetation and atmospheric radiative transfer models (RTMs) in optical remote sensing, discussing methodological underpinnings, synthesizing key applications, and outlining persistent challenges and future research avenues.

Study Configuration

• Spatial Scale: Global to pixel-level, scene-level, and cross-scale transformations (e.g., 100 km x 100 km Sentinel-2 scenes, 10-meter resolution, 30-meter and 300-meter resolution SIF products).
• Temporal Scale: Near real-time to long-term time series analysis.

Methodology and Data

• Models used:
  • Radiative Transfer Models (RTMs): PROSAIL, SCOPE, DART, LESS, 6S, MODTRAN, libRadtran, ACRM, RRTMGP, CRTM, FORUM.
  • Machine Learning Regression Algorithms (MLRAs) / Emulators: Neural Networks (NN), Deep Learning NNs (DLNNs), Gaussian Process Regression (GPR), Random Forests (RF), Support Vector Regression (SVR), Kernel Ridge Regression (KRR), Polynomial Chaos Expansion (PCE), Bayesian Additive Regression Trees (BART), XGBoost, LightGBM, CatBoost, Convolutional NNs (CNNs), Transformers, Physics-Informed NNs (PINNs), Bayesian NNs (BNNs), Generative Adversarial Networks (GANs).
  • Dimensionality Reduction Techniques: Principal Component Analysis (PCA), Singular Value Decomposition (SVD), Autoencoders, Functional PCA (FPCA).
  • Sampling Strategies: Latin Hypercube Sampling (LHS), Sobol sequences, Halton sequences, active learning, adaptive sampling (e.g., AMOGAPE).
• Data sources:
  • Simulated outputs from various RTMs.
  • Empirical observations from airborne (e.g., HyPlant) and satellite (e.g., Sentinel-2, SPOT, MODIS, OCO-2, PRISMA) sensors.
  • LiDAR data for 3D canopy reconstruction.

Main Results

• Emulation accurately replicates RTM outputs with significant computational speed-ups (typically 10^3 to 10^5 times faster), transforming previously intractable problems into feasible solutions.
• MLRAs like GPR, KRR, RF, and (DL)NNs are widely adopted, with GPR providing intrinsic uncertainty quantification and (DL)NNs excelling in high-dimensional and complex data scenarios.
• Spectral dimensionality reduction (e.g., PCA, autoencoders) is critical for efficient handling and reconstruction of high-dimensional spectral data, reducing computational burden and improving generalization.
• Emulators enable efficient global sensitivity analysis (GSA) of RTMs, identifying influential parameters and their interactions in minutes or hours.
• Emulators facilitate fast and scalable synthetic scene generation for sensor-specific applications, satellite mission design, and "what-if" scenario exploration.
• Scene-to-scene emulation allows rapid and accurate transformation of remote sensing products across domains, scales, and sensors (e.g., multispectral to hyperspectral reflectance, airborne to satellite SIF).
• Emulation-based retrieval workflows accelerate the inversion of RTMs for mapping vegetation (e.g., Leaf Area Index, Leaf Chlorophyll Content, Fractional Vegetation Cover) and atmospheric (e.g., aerosol optical depth, water vapor) products from remote sensing data.

Contributions

• Provides a comprehensive review and synthesis of the state-of-the-art in surrogate modeling for optical remote sensing, specifically for vegetation and atmospheric RTMs.
• Systematically discusses the methodological foundations of emulation, including machine learning algorithms, training data sampling, and dimensionality reduction.
• Synthesizes and categorizes the diverse applications of RTM emulation, demonstrating its practical value across various Earth observation tasks.
• Identifies current challenges in generalizability, interpretability, and scalability, and proposes concrete future research directions, including advanced deep learning, multimodal/multitemporal frameworks, and community benchmarking.
• Highlights the transformative potential of emulation in enabling scalable and operational remote sensing workflows for Earth observation.

Funding

Not explicitly mentioned in the provided paper text.

Citation

@article{Verrelst2025Surrogate,
  author = {Verrelst, Jochem and Morata, Miguel and García-Soria, José Luis and Sun, Yilin and Qi, Jianbo and Rivera, Juan Pablo},
  title = {Surrogate Modeling in Optical Remote Sensing: A Review of Emulation for Vegetation and Atmosphere Applications},
  journal = {Preprints.org},
  year = {2025},
  doi = {10.20944/preprints202509.0938.v1},
  url = {https://doi.org/10.20944/preprints202509.0938.v1}
}