LUT-Based Atmospheric Correction

• LUT-based atmospheric correction is a method that converts top-of-atmosphere measurements into bottom-of-atmosphere reflectance by interpolating precomputed radiative transfer outputs.
• It precomputes correction coefficients using RTMs like 6S and libRadtran across varied atmospheric and geometric states, balancing physical fidelity with computational efficiency.
• The technique improves downstream applications such as land-use segmentation by mitigating scattering and gaseous absorption effects in high-resolution Earth observation imagery.

Searching arXiv for recent and foundational papers on LUT-based atmospheric correction and RTM LUT design. Look-Up Table (LUT)-based atmospheric correction is a physically based strategy for converting top-of-atmosphere measurements into bottom-of-atmosphere surface reflectance by precomputing radiative-transfer outputs over a multidimensional state space and then retrieving or interpolating the relevant atmospheric correction coefficients at runtime. In the Earth-observation setting, the purpose is to compensate for atmospheric scattering and gaseous absorption so that spectrally stable, analysis-ready data can be obtained from raw or top-of-atmosphere imagery. In the CARTOSAT-3 MX study, this conversion is treated as a necessary preprocessing step before land-use/land-cover segmentation, with raw “Top-of-Atmosphere (TOA) Digital Number (DN) values” first converted to radiance and then to apparent TOA reflectance, and the final target product being BOA or surface reflectance, also called ARD (Mukherjee et al., 2024).

1. Definition and problem setting

LUT-based atmospheric correction arises from a practical tension between physical fidelity and computational tractability. Physically based atmospheric correction is commonly formulated as inversion of atmospheric radiative transfer equations to recover surface quantities such as reflectance from top-of-atmosphere measurements. RTMs such as 6S, MODTRAN, and libRadtran are preferred because they model absorption and scattering explicitly, but repeated execution over many atmospheric and geometric states is computationally expensive. The operational response is to precompute a multidimensional LUT over selected atmospheric, geometric, and sometimes surface variables, store RTM outputs or derived coefficients at those nodes, and later interpolate between them during processing (Vicent et al., 2020).

In this framework, the LUT does not replace radiative transfer physics; it amortizes it. The standard pattern, described across the cited works, is to evaluate the forward model offline, tabulate atmospheric quantities or correction coefficients over a bounded input domain, and then perform online interpolation or surrogate inference instead of rerunning the RTM for every query. The relevant state variables may include solar/view geometry, aerosol optical depth, water vapour, ozone, elevation, atmospheric profile, aerosol type, and sensor band definitions (Vicent et al., 2020, Mazid et al., 4 May 2026).

For high-resolution optical imagery, the motivation is directly spectral. The CARTOSAT-3 MX study states that raw TOA values are contaminated by atmospheric scattering and gaseous absorption from aerosols, water vapour, and ozone, and that for a very high spatial resolution VIS-NIR sensor these effects can produce haze and band-dependent distortions, especially in the blue band, so the same land-cover class may not have stable spectral signatures unless TOA effects are removed first (Mukherjee et al., 2024).

2. Radiative-transfer formulation and coefficient structure

The core of LUT-based atmospheric correction is a forward radiative-transfer model whose inversion yields surface reflectance. In the CARTOSAT-3 MX formulation, the governing equation is

$$\rho^*\left(\theta_s, \theta_v, \Delta \phi, \lambda\right)= T_g\left(\theta_s, \theta_v, \lambda\right) \times \bigg[\rho_{r a}\left(\theta_s, \theta_v, \Delta \phi, \lambda\right) + \frac{T\left(\theta_s, \lambda\right) T\left(\theta_v, \lambda\right) \rho_s(\lambda)}{1-S(\lambda) \rho_s(\lambda)}\bigg]$$

where $\theta_s$, $\theta_v$, $\Delta \phi$, and $\lambda$ are the sun zenith angle, viewing zenith angle, relative azimuth angle, and wavelength, respectively; $\rho_s$ is BOA surface reflectance; $\rho_{ra}$ is atmospheric path reflectance due to combined aerosol and molecular effects; $T(\theta_s,\lambda)$ and $T(\theta_v,\lambda)$ are downward and upward atmospheric transmittance; $S$ is spherical albedo; and $\theta_s$0 is gaseous transmittance due mainly to ozone and water vapour (Mukherjee et al., 2024).

The apparent reflectance used as the TOA-side input is reported as

$\theta_s$1

with $\theta_s$2 the acquired radiance converted from raw DN value, $\theta_s$3 the extraterrestrial solar irradiance, and $\theta_s$4 the cosine of $\theta_s$5 (Mukherjee et al., 2024).

A closely related compact coefficient form appears in the multi-fidelity coefficient-emulation study, where the coefficient vector is path reflectance, total transmittance, and spherical albedo, and the surface reflectance retrieval equation is given as

$\theta_s$6

That paper emphasizes that LUTs often store precisely these intermediate terms rather than raw radiances, because the coefficients preserve the standard Lambertian correction structure while remaining operationally compact (Mazid et al., 4 May 2026).

This coefficient-centric view clarifies what a LUT usually contains. In the CARTOSAT-3 MX study, the LUT coefficients characterize two main groups of terms: first, $\theta_s$7, which depends on aerosol optical thickness and molecular Rayleigh scattering; second, gaseous absorption and transmittance, which depend primarily on atmospheric water vapour and ozone, along with target surface elevation and acquisition geometry. The paper explicitly states: “The LUT was exploited to interpolate the correction coefficients for the estimated gaseous concentration, aerosol optical thickness (AOT), surface elevation and viewing geometry $\theta_s$8 averaged over the extent for each band present in MX scenes” (Mukherjee et al., 2024).

3. LUT construction, interpolation, and operational workflow

In LUT-based atmospheric correction, the RTM is run offline over a multidimensional domain, after which the online processor queries or interpolates the resulting table. The CARTOSAT-3 MX paper states that “6S was used to simulate Look-Up Tables (LUTs) for the solution vector for a large set of AC coefficients,” rather than downloading a pre-existing table. The LUT is therefore generated offline from 6S runs over a range of atmospheric and geometric conditions and then used during correction to avoid rerunning the radiative transfer model for every image (Mukherjee et al., 2024).

The likely workflow reported for that study is: start with raw DN imagery; convert DN to radiance $\theta_s$9; convert radiance to TOA apparent reflectance using $\theta_v$0; collect or estimate the scene-average atmospheric and geometric inputs needed by the LUT; interpolate the corresponding coefficients from the precomputed LUT; and then invert the forward model to recover $\theta_v$1, band by band. The paper does not provide the DN-to-radiance calibration formula, the explicit inversion algebra for solving the forward equation for $\theta_v$2, or a formal interpolation equation for the LUT search, so those elements are unreported even though the operational sequence is conceptually clear (Mukherjee et al., 2024).

The general interpolation problem has been studied directly. GALGA formulates LUT construction as an optimal-design problem for costly RTMs, with linear interpolation over a Delaunay triangulation:

$\theta_v$3

where $\theta_v$4 are barycentric coordinates of the query point with respect to the enclosing simplex (Vicent et al., 2020). That work addresses node placement rather than atmospheric correction inversion itself, but it is directly relevant because interpolation accuracy depends strongly on LUT geometry. Relative to a pseudo-random homogeneous distribution of LUT nodes, GALGA reports reducing the LUT size by $\theta_v$5 and the maximum interpolation relative errors by 0.5% (Vicent et al., 2020).

A broader comparison between interpolation and learned emulation reaches an even stronger conclusion for sparse RTM LUTs. For MODTRAN-based TOA radiance reconstruction, nearest-neighbour, inverse distance weighting, and piece-wise linear interpolation were compared against Gaussian process regression, kernel ridge regression, and neural networks. In all experiments, the emulation methods produced more accurate output spectra than classical interpolation methods, with GPR performing more than 10 times better than the best interpolation baseline in the MODTRAN case while remaining much faster than piece-wise linear interpolation (Vicent et al., 2020). This suggests that, within LUT-based atmospheric correction, the online query mechanism itself is now an active design variable rather than a fixed interpolation choice.

4. Sensor adaptation and the CARTOSAT-3 MX implementation

The most explicit Earth-observation implementation in the cited set is the CARTOSAT-3 MX workflow for high-resolution urban imagery. The method is sensor-adapted in the sense that interpolation is performed “for each band present in MX scenes,” and the spectral support is reported as Blue $\theta_v$6–$\theta_v$7m, Green $\theta_v$8–$\theta_v$9m, Red $\Delta \phi$0–$\Delta \phi$1m, and NIR $\Delta \phi$2–$\Delta \phi$3m (Mukherjee et al., 2024). The safest reading is that the LUT is band-specific for MX; however, whether the 6S simulations used full sensor spectral response functions or nominal band ranges is not reported (Mukherjee et al., 2024).

The imagery spans six Indian cities. Training and development imagery covers Bangalore, Mumbai, Pune, Varanasi, and Delhi, while the main atmospheric correction result and evaluation example emphasized in the paper is a subset over Hyderabad acquired on March 12, 2023. The data are described as “very high-resolution (around $\Delta \phi$4) multispectral,” and correction appears to be applied on a scene/subscene basis rather than via fully per-pixel atmospheric retrieval, because the geometry terms $\Delta \phi$5 were “averaged over the extent” and the LUT coefficients were interpolated for the estimated atmospheric state and geometry for each band (Mukherjee et al., 2024).

The atmospheric components being compensated are clearly identified: atmospheric path reflectance $\Delta \phi$6, gaseous transmittance $\Delta \phi$7, upward and downward transmittances, and spherical albedo $\Delta \phi$8. By contrast, adjacency effects are not corrected, even though they are discussed as a challenge in the literature, and the authors explicitly state that “our approach currently lacks a calculation for water surface glint” (Mukherjee et al., 2024). These omissions define the boundaries of the method as presented.

The paper also leaves several replication-critical details unspecified: the DN calibration coefficients, the exact LUT grid design, the source of AOT, water-vapour, ozone, and elevation inputs, the atmospheric profile used in 6S, whether relative azimuth was included in interpolation, whether spectral response functions were used, the interpolation method itself, and the exact inversion formula used to recover $\Delta \phi$9 (Mukherjee et al., 2024). For reproducibility, these missing elements are as important as the reported forward model.

5. Validation, downstream use, and empirical behavior

The empirical validation in the CARTOSAT-3 MX study is primarily qualitative and task-driven. The corrected Hyderabad image is visually less hazy than the TOA reflectance image, and the paper states that BOA reflectance values are reduced after correction in a spectrally dependent way, “with the blue band being mostly affected by the ARD procedure,” which is interpreted as haze removal from molecular scattering (Mukherjee et al., 2024). Figure 1 is summarized as showing TOA reflectance, atmospherically corrected BOA reflectance, an intercomparison of TOA and BOA reflectance for four LULC classes, and LUT-derived mean path reflectance with $\lambda$0 uncertainty envelopes (Mukherjee et al., 2024).

The main quantitative evidence is downstream segmentation performance. The corrected image, described as ARD, is used as input to a supervised DeepLabv3+ model and a semi-supervised Cross Pseudo Supervision (CPS) model for Trees, Buildings, Water, and Roads on a $\lambda$1 subsection of the Hyderabad image. Under sparse labels, recall is the reported metric (Mukherjee et al., 2024).

For the semi-supervised CPS model, recall improves from 78.5390 to 93.114 for Trees, from 65.1801 to 74.3087 for Buildings, from 66.0376 to 86.8159 for Water, and from 58.4839 to 62.0745 for Roads when switching from Raw to ARD. For the supervised DeepLabv3+ model, recall changes from 16.6227 to 50.9215 for Trees, from 53.2109 to 57.8633 for Buildings, from 70.0857 to 50.3246 for Water, and from 0.0410 to 40.0537 for Roads (Mukherjee et al., 2024). The broad conclusion is therefore not universal per-class improvement, but stability and strong gains in the semi-supervised setting, especially under sparse labeling (Mukherjee et al., 2024).

This suggests that LUT-based atmospheric correction can matter operationally not only because it produces BOA reflectance, but because it regularizes spectral inputs for downstream learning. A plausible implication is that pseudo-labeling methods such as CPS are particularly sensitive to atmospheric inconsistency, since input spectral biases can be reinforced during self-training. The paper itself does not isolate whether the gains come from DN-to-reflectance normalization alone, from the full 6S-based atmospheric correction, or from other preprocessing choices; it only compares “ARD” versus “Raw” (Mukherjee et al., 2024).

6. Design trade-offs, alternatives, and contemporary extensions

The recent literature around LUT-based atmospheric correction has increasingly focused on the LUT as an object of optimization. One line of work improves node placement. GALGA uses an acquisition function combining a geometry term and a density term to adaptively choose where RTM evaluations should be performed, rather than relying on regular grids or homogeneous pseudo-random sampling. Its practical significance is that fewer RTM runs are needed to reach a specified interpolation-error target, which directly reduces offline generation cost and storage (Vicent et al., 2020).

A second line of work replaces interpolation with emulation. The emulation study using MODTRAN concludes that sparse LUTs plus GPR or KRR can reconstruct TOA radiance spectra more accurately than nearest-neighbour, inverse distance weighting, or piece-wise linear interpolation, with better runtime characteristics in the high-dimensional scattered setting (Vicent et al., 2020). In this view, the LUT is no longer merely a table to be interpolated; it becomes a training design for a surrogate model.

A third line retains the coefficient-based physics but upgrades the fidelity of coefficient generation. The pKANrtm framework uses paired 6S and libRadtran simulations, sensor-band spectral-response-function-aware coefficient generation, and residual learning in coefficient space to emulate high-fidelity path reflectance, total transmittance, and spherical albedo for 13 Sentinel-2 bands (Mazid et al., 4 May 2026). The reported single-sample runtimes are $\lambda$2 ms/sample for libRadtran CPU, $\lambda$3 ms/sample for 6S CPU, $\lambda$4 ms/sample for pKANrtm GPU single sample, and $\lambda$5 ms/sample amortized for pKANrtm GPU batch 128, with the model achieving RMSE $\lambda$6, MAE $\lambda$7, $\lambda$8, and SMAPE $\lambda$9 on the standard split (Mazid et al., 4 May 2026).

These developments do not negate the classical LUT formulation. Rather, they shift emphasis from a static dense grid toward adaptive design, surrogate querying, and coefficient-level emulation. The common thread is that the physically interpretable atmospheric correction equation is preserved, while the expensive RTM evaluation layer is compressed, redesigned, or statistically emulated (Vicent et al., 2020, Mazid et al., 4 May 2026).

7. Misconceptions, limitations, and scope

A common misconception is that LUT-based atmospheric correction is intrinsically synonymous with a fixed, coarse, or purely empirical correction. The cited papers indicate the opposite: LUT workflows are physically grounded because they are derived from RTM simulations, and their limitations are usually traceable to LUT design, interpolation error, incomplete atmospheric state specification, or unmodeled phenomena rather than to the LUT concept itself (Vicent et al., 2020, Mukherjee et al., 2024).

Another misconception is that atmospheric correction quality can be inferred directly from downstream task improvement. The CARTOSAT-3 MX study does not validate BOA reflectance against in situ spectroradiometric measurements, field targets, RadCalNet, or cross-sensor reflectance, and it reports no quantitative reflectance error metrics such as RMSE, bias, or spectral angle (Mukherjee et al., 2024). Its evidence for correction quality is mainly visual and indirect, supplemented by downstream segmentation gains. That makes it a task-oriented demonstration rather than a formal reflectance-validation study.

A further point of clarification concerns domain scope. The paper on on-sky closed-loop correction of atmospheric dispersion also uses a LUT, but in a different sense: an Atmospheric Dispersion Compensator is commanded from a look-up table mapping telescope elevation to prism settings, and the paper shows that LUT-only open-loop correction is insufficient for sub-milliarcsecond residuals in high-contrast coronagraphy and astrometry (Pathak et al., 2017). This is not Earth-observation reflectance correction, but it illustrates a general systems principle: a LUT is often an excellent feed-forward prior, yet not necessarily sufficient when residuals must be controlled in the final measurement space (Pathak et al., 2017). This suggests an analogous caution for remote sensing: LUT-based preprocessing can be physically principled while still requiring careful attention to residual errors, missing physics, and application-specific validation.

Within optical Earth observation, the main limitations documented in the CARTOSAT-3 MX implementation are explicit: adjacency effects are not corrected; sunglint for water is absent; atmospheric inputs are only described as “estimated”; geometry is averaged over the scene extent; and implementation details critical for exact replication are unreported (Mukherjee et al., 2024). In that sense, LUT-based atmospheric correction is best understood not as a single algorithm, but as a family of RTM-driven inversion workflows whose performance depends on coefficient selection, LUT state-space design, interpolation or emulation strategy, sensor adaptation, and the completeness of the modeled atmospheric terms.