Radiative transfer models (RTMs) are foundational to optical remote sensing for simulating vegetation and atmospheric properties. However, their significant computational cost, especially for 3D RTMs and large-scale applications, severely limits their utility. Emulation, or surrogate modeling, has emerged as a highly effective strategy, accurately and efficiently replicating RTM outputs. This review comprehensively surveys recent developments in emulating vegetation and atmospheric RTMs. We discuss the methodological underpinnings, including suitable machine learning regression algorithms (MLRAs), effective training sampling strategies (e.g., Latin Hypercube Sampling, active learning), and spectral dimensionality reduction (DR) methods (e.g., PCA, autoencoders). Emulators commonly achieve ${10}^2-{10}^6\times$ per-evaluation acceleration, but accuracy–efficiency trade-offs remain inherently context-dependent, governed by the MLRA design and the coverage/quality of training data. DR consistently shifts this trade-off toward lower cost at comparable accuracy, positioning latent-space training as a pragmatic choice for hyperspectral applications. We synthesize key emulation applications such as global sensitivity analysis, synthetic scene generation, scene-to-scene translation (e.g., multispectral-to-hyperspectral), and retrieval of geophysical variables using remote sensing data. The paper concludes by outlining persistent challenges in generalizability, interpretability, and scalability, while also proposing future research avenues: investigating advanced deep learning algorithms (e.g., physics-informed and explainable architectures), developing multimodal/multitemporal frameworks, and establishing community benchmarks, tools and libraries. Emulation ultimately empowers remote sensing workflows with unparalleled scalability, transforming previously unmanageable tasks into viable solutions for operational Earth observation applications.

Radiative transfer models (RTMs) are foundational to optical remote sensing for simulating vegetation and atmospheric properties. However, their significant computational cost, especially for 3D RTMs and large-scale applications, severely limits their utility. Emulation, or surrogate modeling, has emerged as a highly effective strategy, accurately and efficiently replicating RTM outputs. This review comprehensively surveys recent developments in emulating vegetation and atmospheric RTMs. We discuss the methodological underpinnings, including suitable machine learning regression algorithms, effective training sampling strategies (e.g., Latin Hypercube Sampling, active learning), and spectral dimensionality reduction methods (e.g., PCA, autoencoders). We then synthesize key emulation applications such as global sensitivity analysis, synthetic scene generation, scene-to-scene translation (e.g., multispectral-to-hyperspectral), and retrieval of geophysical variables using remote sensing data. The paper concludes by outlining persistent challenges in generalizability, interpretability, and scalability, while also proposing future research avenues: investigating advanced deep learning algorithms (e.g., physics-informed and explainable architectures), developing multimodal/multitemporal frameworks, and establishing community benchmarks, tools and libraries. Emulation ultimately empowers remote sensing workflows with unparalleled scalability, transforming previously unmanageable tasks into viable solutions for operational Earth observation applications.