Radiometric Consistency in Remote Sensing

• Radiometric consistency is the preservation of physically meaningful measurements across different sensors, conditions, and processing techniques.
• It involves calibration and normalization methods in satellite imagery, planetary mosaicking, CT imaging, and inverse rendering to maintain data comparability.
• Robust techniques such as vicarious and relative calibration improve metrics like PSNR, SSIM, and radiomic reproducibility for reliable analysis.

Radiometric consistency denotes the condition under which measured image values, detector outputs, or derived quantitative features remain physically comparable across acquisitions, sensors, viewing conditions, or processing stages. In satellite remote sensing, it is the requirement that image values correspond as closely as possible to the actual physical radiation associated with the Earth’s surface rather than to variable atmospheric attenuation, illumination geometry, or sensor-view effects (Delgado-Correal et al., 2012). In planetary mosaicking, the same term is used more operationally to describe adjacent tiles with compatible brightness, contrast, and tonal response, so that mosaics are illumination-consistent and free of obvious seam lines (Singh et al., 28 Apr 2026). Across other fields, the concept extends to preserving per-pixel temperature meaning in radiometric thermal rasters, ensuring reproducibility of quantitative CT features across acquisition settings, or forcing inverse-rendered radiance fields to satisfy the rendering equation rather than merely fit images (Habibpour et al., 25 Jun 2026, Erdal et al., 2019, Hadadan et al., 2023).

1. Conceptual scope and domain-specific meanings

Radiometric consistency is not a single protocol but a family of requirements linking recorded signals to the physical process that generated them. In optical Earth observation, passive sensors record radiation coming from both the surface and the atmosphere, so the same site can produce different pixel values across dates even when the surface has not changed. Consistency therefore depends on transforming image values toward radiance, reflectance, and ideally surface reflectance, while accounting for atmospheric absorption, scattering, solar geometry, and directional reflectance effects (Delgado-Correal et al., 2012).

In relative normalization problems, especially when the goal is change detection, object classification, or map mosaicking, radiometric consistency is often defined more pragmatically. The central requirement is that unchanged terrain should remain comparable across scenes, while real changes, clouds, or fog should not dominate the normalization model. That is the rationale for robust relative radiometric normalization methods that estimate a no-change set and learn the mapping from those pixels rather than from the whole overlap region (Liu et al., 2021).

In camera pipelines and image formation, radiometric consistency often means restoring a linear or approximately linear relation between recorded values and scene radiance. JPEG images violate that relation because they have passed through color correction, gamut mapping, and tone mapping, so radiometric calibration becomes the task of recovering an approximate mapping between rendered JPEG values and raw sensor values (Gong et al., 2017). In camera-display communication, the analogous problem is the camera-display transfer function, where displayed intensity and captured intensity are not radiometrically consistent because both the display and the camera are nonlinear and the display emittance is view-dependent (Yuan et al., 2015).

In physically based vision and inverse rendering, the notion becomes stricter still. A scene is radiometrically consistent when one coherent set of reflectance, illumination, geometry, and radiance explains all observations under the rendering equation, including shadows, interreflection, and view dependence (Lombardi et al., 2016, Han et al., 2 Mar 2026). In quantitative imaging outside visible remote sensing, the same principle appears as reproducibility: CT-derived density, histogram, and texture features are useful only when they do not change materially under routine variations in dose, reconstruction kernel, or slice thickness (Erdal et al., 2019).

2. Physical basis in optical remote sensing

The physical foundations of radiometric consistency in remote sensing begin with the Sun–atmosphere–surface–sensor system. The incoming illumination has a structured spectral dependence, treated in conceptual form through Planck’s law for blackbody emission, and the detector does not record ground properties directly but directional spectral radiance that has already been modified by the atmosphere and the surface (Delgado-Correal et al., 2012). In this framework, the sensor-reaching radiance is shaped by both multiplicative attenuation and additive path contributions.

The atmospheric transfer relation is written as

$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f$

with

$\gamma_A = \gamma_a + \gamma_s,$

where $L_f$ is spectral radiance in direction $(\theta,\phi)$, $J_f$ is radiance scattered into that direction, $z$ is path length, and $\gamma_a$ and $\gamma_s$ are the absorption and scattering coefficients (Delgado-Correal et al., 2012). This expression makes clear why two images of the same target can differ radiometrically when aerosol load, molecular scattering, cloud presence, or gas absorption differ.

At the surface, radiometric consistency also depends on how incident energy is partitioned:

$E_I = E_R + E_A + E_T.$

The reflected component is therefore not arbitrary image brightness but the remainder after absorption and transmission, and different materials generate different spectral signatures because their partitioning varies with wavelength (Delgado-Correal et al., 2012). The same review emphasizes that reflectance is not purely intrinsic; it is directional. Irradiance from the illumination direction and radiance toward the sensor are linked through the bidirectional reflectance distribution function,

$$BRDF = \frac{dL_r(\theta_r,\phi_r;\theta_i,\phi_i)}{dE_i},$$

so even perfect sensor calibration and atmospheric correction do not guarantee inter-image equality when sun angle and view angle differ (Delgado-Correal et al., 2012).

For that reason, remote-sensing radiometric consistency is typically improved in stages. A standard top-of-atmosphere reflectance normalization is

$\gamma_A = \gamma_a + \gamma_s,$0

which removes variation due to Earth–Sun distance and solar zenith angle (Delgado-Correal et al., 2012). A further surface-reflectance step must model atmospheric transmission and path radiance. The same physical logic motivates in-situ calibration strategies using field spectroscopy. In one scene-specific vicarious calibration study, a spectroradiometer and NOAA-18 AVHRR visible-band data were linked through an instrumental transfer function and a total atmospheric attenuation factor, yielding

$\gamma_A = \gamma_a + \gamma_s,$1

over 430–830 nm for the selected pixel (Delgado-Correal et al., 2012). This result was interpreted as the total attenuation of integrated radiation intensity due to the atmosphere and viewing geometry.

3. Calibration and normalization methodologies

A wide range of methodologies pursue radiometric consistency, but they differ in whether the target is absolute physical calibration, relative temporal harmonization, or operational tonal normalization.

A scene-specific in-situ calibration workflow couples field spectroscopy to orbital imagery through a sensor-response transfer function and an atmospheric attenuation estimate (Delgado-Correal et al., 2012). This is closest to vicarious calibration in the classical remote-sensing sense. By contrast, multi-temporal UAV monitoring often uses relative scene-to-scene calibration. In coal spoil monitoring, the proposed method was a relative empirical line calibration with invariant targets, denoted ELC-IT, where unchanged spoil piles in overlapping orthomosaics supplied the reference points. For each band, the target image was mapped to the reference with a through-origin linear model,

$\gamma_A = \gamma_a + \gamma_s,$2

and calibration was applied sequentially across dates (Thiruchittampalam et al., 2024). The study reported statistically significant regressions for all date pairs and bands, and the best calibrated classification pipeline reached 90.7% overall accuracy with ensemble (subspace discriminant), compared with 83.0% overall accuracy for the best uncalibrated pipeline (Thiruchittampalam et al., 2024).

Relative radiometric normalization can also be formulated probabilistically. In latent change-noise modeling, the observation model is

$\gamma_A = \gamma_a + \gamma_s,$3

with latent no-change and change components, and the normalization mapping is learned jointly with mixture parameters through EM (Liu et al., 2021). The best-performing variant, HM-RRN-MoG, combines a Gaussian mixture noise model with histogram matching, using the prior that no-change pixels should exhibit small residual noise after normalization while changed pixels should exhibit large residual noise. This formulation yields a robust no-change set and a no-change-set RMSE that directly evaluates radiometric consistency on invariant terrain (Liu et al., 2021).

Exposure control is another calibration variable rather than a mere acquisition convenience. In UAV multispectral phenotyping, fixed exposure was shown to improve radiometric fidelity relative to auto-exposure because exposure and gain changes varied by band and by scene content under the default camera behavior (Swaminathan et al., 2024). The raw-to-radiance step was modeled as

$\gamma_A = \gamma_a + \gamma_s,$4

with vignette correction $\gamma_A = \gamma_a + \gamma_s,$5, gain $\gamma_A = \gamma_a + \gamma_s,$6, black level $\gamma_A = \gamma_a + \gamma_s,$7, and exposure time $\gamma_A = \gamma_a + \gamma_s,$8 (Swaminathan et al., 2024). After panel correction and object-based empirical line calibration, fixed-exposure orthomosaics were closer to ASD ground truth than auto-exposure orthomosaics, and the study identified band-specific ideal exposure ranges within which MAPE < 5% for blue, green, red, and NIR, and < 7% for red edge in the abstract’s summary (Swaminathan et al., 2024).

In consumer-camera contexts, radiometric consistency often requires approximate inversion of a camera rendering pipeline rather than atmospheric correction. A compact model writes the forward mapping as

$\gamma_A = \gamma_a + \gamma_s,$9

where $L_f$0 is a $L_f$1 color correction matrix, $L_f$2 is gamut mapping, and $L_f$3 is tone mapping (Gong et al., 2017). The key observation is that rank order is mostly preserved after color correction, making it possible to estimate $L_f$4 from pairwise ranking constraints and then recover monotonic tone curves and a residual LUT. This produces a radiometrically calibrated approximation between raw and JPEG using only a single raw/JPEG pair (Gong et al., 2017).

4. Cross-sensor and cross-platform consistency

Radiometric consistency is often required where heterogeneous sensors or channels must be placed on a common scale. The practical formulations differ, but the underlying need is the same: values should remain interpretable after sensor fusion.

In lunar mosaicking, the problem is visible seam artifacts and tonal discontinuities between Chandrayaan-2 TMC and SELENE/Kaguya tiles. The proposed framework used a conditional generative adversarial network with a U-Net generator and PatchGAN discriminator to map a conventionally mosaicked TMC+SELENE image $L_f$5 to a normalized output $L_f$6 aligned with an LROC WAC reference domain (Singh et al., 28 Apr 2026). The training objective combined an adversarial term with an $L_f$7 reconstruction term,

$L_f$8

with $L_f$9, and large mosaics were reconstructed from overlapping patches by weighted blending,

$(\theta,\phi)$0

The best reported checkpoint reached PSNR $(\theta,\phi)$1 dB and SSIM $(\theta,\phi)$2 at epoch 125, with reduced seam artifacts and improved tonal uniformity (Singh et al., 28 Apr 2026).

A related cross-sensor problem appears in super-resolution of Sentinel-2 using PlanetScope supervision. Because PlanetScope is used as the high-resolution target, a naïve model tends to drift toward PlanetScope color and contrast even though Sentinel-2 Level-2A products are treated as the more radiometrically trustworthy source. The proposed consistency term therefore constrains the super-resolved Sentinel-2-like output so that, once downsampled, it matches the source Sentinel-2 reference frame, while a separate color-matching head adapts the output toward PlanetScope for the super-resolution loss (Razzak et al., 2021). This design was intended to preserve Sentinel-2 radiometric properties while still borrowing spatial detail from multi-image fusion and cross-sensor supervision.

Inflight calibration of multispectral space instruments illustrates the same issue at detector level. For New Horizons MVIC, one calibration path compares observed stellar fluxes to modeled stellar fluxes to derive detector-specific adjustment factors, while a second path uses Charon color ratios to calibrate Blue, NIR, and CH$(\theta,\phi)$3 relative to Red (Howett et al., 2016). The two methods agree to better than 7%, which the paper treats as strong validation. Detector-specific stellar calibration factors include 1.21 ± 0.01 for Red, 1.00 ± 0.01 for Blue, 1.32 ± 0.01 for NIR, and 1.46 ± 0.02 for CH$(\theta,\phi)$4 (Howett et al., 2016). Similarly, the revised Hinode/EIS calibration uses line-ratio consistency across its SW and LW channels to show that the LW channel degraded strongly with time, while the revised first-order correction brings the main cross-channel ratios to within a relative 20% and removes long-standing discrepancies of factors larger than two for several key ions (Zanna, 2012).

Mission-scale radiometric models generalize these ideas from detector response to observing programs. ArielRad uses a common forward model for all channels and targets, with wavelength-dependent throughput, detector response, background terms, and a unified noise formalism (Mugnai et al., 2020). In that setting, radiometric consistency means that every target and observing tier is evaluated under the same signal and uncertainty model, enabling internally consistent performance predictions and the conclusion that Ariel is generally photon-noise dominated in the infrared (Mugnai et al., 2020).

5. Quantitative measurement fields: thermal wildfire monitoring and CT

In thermal wildfire monitoring, radiometric consistency is tied directly to whether thermal labels are grounded in per-pixel temperature rather than in a qualitative color palette. FlameVQA defines the radiometric thermal TIFF, not the colorized thermal JPEG, as the authoritative thermal signal. The TIFF is described as a single-band raster with per-pixel temperature values, and the benchmark uses it to compute temperature summaries, validate temperature-critical labels, and define deterministic hotspot masks (Habibpour et al., 25 Jun 2026). The central hotspot rule is

$(\theta,\phi)$5

with $(\theta,\phi)$6 interpreted as per-pixel temperature in $(\theta,\phi)$7 and $(\theta,\phi)$8 a fixed physical threshold (Habibpour et al., 25 Jun 2026). The paper is explicit that supervision is grounded in absolute temperatures rather than normalized heat patterns.

WildFireVQA develops the same principle at larger scale. Each sample contains an RGB image, a color-mapped thermal visualization, and a radiometric thermal TIFF, and deterministic thermal labels are derived directly from the temperature field $(\theta,\phi)$9 (Habibpour et al., 22 Apr 2026). One canonical hotspot mask is

$J_f$0

with connected components filtered by altitude-aware ground-sampling-distance rules to suppress isolated sensor artifacts and numerical noise (Habibpour et al., 22 Apr 2026). The benchmark further derives hotspot area, spacing, intensity consistency, and percent-above-threshold labels from the TIFF, while using intra-frame and inter-frame consistency checks to reduce annotation contradictions (Habibpour et al., 22 Apr 2026). In both wildfire benchmarks, radiometric consistency is therefore a property of benchmark construction and label validity rather than of a new sensor-calibration pipeline.

In CT radiomics, the corresponding concept is reproducibility under acquisition and reconstruction changes. Lung nodule scans reconstructed under 320 combinations of dose, kernel, and slice thickness showed that volumetric reproducibility decreases as thickness increases, whereas reproducibility of histogram- and texture-based features improves (Erdal et al., 2019). The paper defines compatible pairs using a $J_f$1-statistic threshold,

$J_f$2

and summarizes radiomic compatibility by

$J_f$3

The strongest message is that quantitative features are not inherently portable across dose, kernel, and thickness choices; radiometric or radiomic consistency requires balanced protocol standardization (Erdal et al., 2019).

6. Scene-level and physically based formulations

In inverse rendering and physically based computer vision, radiometric consistency is no longer merely inter-image comparability. It becomes the requirement that radiance itself satisfy the scene’s light-transport law.

Radiometric Scene Decomposition formulates the problem in a Bayesian framework,

$J_f$4

where all images must be explained by a single reflectance, illumination, and geometry model (Lombardi et al., 2016). The method uses the rendering equation with direct and indirect illumination, non-Lambertian BRDFs, 3D material segmentation, diffuse texture separation, and geometry refinement. In this context, radiometric consistency means that shadows, interreflections, and specularities are explained by the same scene variables rather than absorbed arbitrarily into texture or albedo (Lombardi et al., 2016).

A neural radiometric prior provides a more explicit fixed-point view. The rendering equation is written as

$J_f$5

and the residual prior penalizes the rendering-equation defect of a neural radiance function $J_f$6,

$J_f$7

This prior is designed to enforce physical transport consistency beyond the observed pixels and reduce the need to differentiate long Monte Carlo path integrals directly (Hadadan et al., 2023).

Radiometrically Consistent Gaussian Surfels adopts the same principle for Gaussian-splatting inverse rendering. The learned outgoing radiance $J_f$8 is constrained to match its physically based counterpart,

$J_f$9

with loss

$z$0

The purpose is to supervise radiance along unobserved directions, where ordinary novel-view supervision is absent but indirect illumination depends critically on correctness (Han et al., 2 Mar 2026). Ablations reported drops from 37.86 to 35.82 in NVS PSNR and from 32.09 to 31.69 in relight PSNR when $z$1, supporting the claim that the radiometric-consistency term improves interreflection modeling and disentanglement (Han et al., 2 Mar 2026).

The same theme appears in active imaging. In camera-display communication, the captured signal is governed by a camera-display transfer function rather than simple surface reflectance. The method therefore models an inverse radiometric response

$z$2

and embeds calibration directly into a convex message-decoding formulation via a physics-based feature lift (Yuan et al., 2015). In raw/JPEG calibration, rank preservation after color correction supplies a similarly compact route back toward a radiometrically meaningful raw-like domain (Gong et al., 2017).

7. Evaluation criteria, limitations, and future directions

Radiometric consistency is usually assessed indirectly, through invariants that should not change if calibration is correct. In spectroscopy, this means branching ratios, weakly density-sensitive ratios, and quiet-Sun radiances; in multitemporal normalization, it means the error on inferred no-change pixels; in mosaicking and super-resolution, it often means PSNR, SSIM, histograms, and visible seam suppression; in wildfire thermal benchmarks, it means agreement between labels and radiometric TIFF-derived rules; and in inverse rendering, it means novel-view fidelity, relighting accuracy, or residual agreement with the rendering equation (Zanna, 2012, Liu et al., 2021, Singh et al., 28 Apr 2026, Habibpour et al., 22 Apr 2026, Hadadan et al., 2023).

These evaluations also expose the main limitations. Remote-sensing calibration remains vulnerable to imperfect atmospheric models, especially in tropical conditions where standard models may be poorly matched to the true atmosphere (Delgado-Correal et al., 2012). Relative normalization methods depend on sufficient geometric alignment and on the existence of a valid no-change set (Liu et al., 2021). UAV multispectral calibration depends strongly on exposure control; once bright targets saturate beyond the ideal exposure range, post hoc correction cannot recover the lost information (Swaminathan et al., 2024). Cross-sensor deep normalization methods may reduce seams and preserve structure, but can still omit critical implementation details such as registration, resampling, or baseline quantification (Singh et al., 28 Apr 2026). Benchmark designs that rely on radiometric TIFFs can preserve physical meaning while still leaving upstream emissivity, atmospheric correction, and calibration drift undocumented (Habibpour et al., 25 Jun 2026, Habibpour et al., 22 Apr 2026). Inverse-rendering approaches that impose radiometric priors remain approximate when geometry is fixed, gradients are biased, or specular transport is not modeled directly (Hadadan et al., 2023, Han et al., 2 Mar 2026).

A recurrent direction in the literature is hybridization: combining physics-based constraints with learned models rather than treating them as alternatives. Lunar mosaicking identifies physics-informed learning strategies that incorporate lunar photometric models as future work (Singh et al., 28 Apr 2026). Multitemporal UAV monitoring recommends integrating irradiance sensors with relative empirical-line workflows (Thiruchittampalam et al., 2024). Wildfire VQA benchmarks point toward models that directly consume radiometric TIFF information rather than relying on text retrieval of summary statistics (Habibpour et al., 22 Apr 2026). Quantum radiometric calibration pushes the concept further by calibrating photodiode efficiency in situ at the actual sideband frequency of use, using squeezed vacuum states and the Heisenberg relation rather than conventional power standards alone (Albers et al., 16 Dec 2025).

Taken together, these works show that radiometric consistency is best understood as a constraint of physical comparability. Whether the object is a satellite pixel, a lunar mosaic tile, a multispectral orthomosaic, a thermal hotspot mask, a CT texture feature, or a learned radiance field, the central question is the same: do the recorded values preserve the physical quantity or relation that downstream interpretation assumes? The answer depends on calibration, normalization, geometry, instrument behavior, and model choice, but the governing principle remains the preservation of physically meaningful measurement across transformations (Delgado-Correal et al., 2012).