CodeCompass: Radiometric Consistency Insights

• CodeCompass is a framework that defines radiometric consistency as the invariant link between measured signals and physical properties across multiple domains.
• It integrates methodologies like empirical line calibration, nonlinear mapping, and physics-based correction to overcome issues from atmospheric and sensor variations.
• By quantifying performance with metrics such as PSNR, SSIM, and calibration accuracy, CodeCompass provides actionable insights for reliable imaging and analysis.

Radiometric consistency is the condition under which measured values remain tied to the physical radiation, reflectance, temperature field, or detector response they are intended to represent, rather than drifting because of atmosphere, illumination geometry, sensor behavior, reconstruction settings, or downstream rendering. Across the literature, the term has domain-specific meanings: in satellite remote sensing it denotes comparability of radiance or reflectance across dates, scenes, and sensors; in planetary mosaicking it denotes seam-free tonal harmonization relative to a stable reference domain; in inverse rendering it denotes agreement between a learned radiance field and the rendering equation; in thermal benchmarks it denotes labels grounded in per-pixel temperature-bearing rasters rather than palette images; and in meteoritic chronology it denotes concordance among independent radiometric clocks (Delgado-Correal et al., 2012, Singh et al., 28 Apr 2026, Hadadan et al., 2023, Habibpour et al., 22 Apr 2026, Desch et al., 2022).

Domain	Operational meaning of radiometric consistency	Representative sources
Earth and planetary imaging	Comparable radiance/reflectance or harmonized tonal response across acquisitions	(Delgado-Correal et al., 2012, Liu et al., 2021, Singh et al., 28 Apr 2026)
Instrument calibration	Stable conversion from counts to physical units across channels and time	(Howett et al., 2016, Zanna, 2012, Albers et al., 16 Dec 2025)
Computer vision and inverse rendering	One physically compatible explanation of appearance, illumination, and transport	(Lombardi et al., 2016, Hadadan et al., 2023, Han et al., 2 Mar 2026)
Thermal and quantitative benchmarks	Use of radiometric TIFFs or quantitative measurements as authoritative supervision	(Habibpour et al., 25 Jun 2026, Habibpour et al., 22 Apr 2026)
Quantitative analysis outside optical imaging	Reproducible quantitative features or concordant chronometers across protocols	(Erdal et al., 2019, Desch et al., 2022)

1. Physical basis: radiation, atmosphere, surface, and directionality

In optical remote sensing, radiometric consistency begins with the fact that a pixel value is not a direct ground property. Passive sensors record radiation reflected from the surface and modified by the atmosphere, so equal surfaces can produce unequal image values when atmospheric absorption, scattering, illumination geometry, or viewing geometry differ. The physical chain is therefore Sun–atmosphere–surface–sensor, not surface alone (Delgado-Correal et al., 2012).

The physical description starts from wavelength-dependent solar illumination. The review of satellite-image calibration gives Planck’s law for the spectral distribution of blackbody emission,

$$B_\lambda = \frac{2hc^2}{\lambda^5\left(e^{\frac{hc}{\lambda k_B T}-1}\right)},$$

to emphasize that incoming irradiance is spectrally structured before any interaction with the Earth system. Atmospheric propagation is then described by

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

so the detector receives not only attenuated surface radiance but also additive scattered radiance. This is why calibration must model both absorption and scattering, and why band dependence matters for gases such as $\mathrm{O_2, N_2, O_3, CO_2, H_2O}$ (Delgado-Correal et al., 2012).

At the surface, the same review expresses the partition of incident energy as

$E_I = E_R + E_A + E_T,$

and then moves from reflected energy to reflectance via

$$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$

It also makes directional dependence explicit through the bidirectional reflectance distribution function,

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

showing that radiometric consistency is not guaranteed even after sensor calibration and atmospheric correction if the surface is anisotropic and sun–sensor geometry changes (Delgado-Correal et al., 2012).

An in-situ calibration study makes the same point operational. Using a field spectroradiometer and NOAA-18 AVHRR visible-band imagery, it models the spectroradiometer-to-satellite relationship as an instrumental transfer function,

$$H_{\text{instrumental}(\lambda)} = \frac{F_{\text{AVHRR spectral response}(\lambda)}}{F_{\text{spectroradiometer spectral response}(\lambda)}},$$

then compares the simulated near-ground AVHRR response with the real orbital pixel to estimate a total atmospheric attenuation factor. For the case study, over $430$–$830$ nm, the reported value was

$$H_{\mathrm{Total}(\lambda)\,\mathrm{Atmosphere}} = 1.435 \times 10^{-3},$$

under the assumption of a spatially homogeneous atmosphere (Delgado-Correal et al., 2012).

2. Earth-observation workflows: temporal harmonization, invariant targets, and exposure control

In multi-temporal Earth observation, radiometric consistency is often treated as scene-to-scene comparability rather than absolute reflectance recovery. A UAV coal-spoil study states that raw RGB values are not temporally comparable by default because of sensor noise, atmospheric scattering and absorption, variations in sun parameters, and variable characteristics of the sensed object over time. Its solution is a relative empirical line calibration with invariant targets, termed ELC-IT, using pseudo-invariant spoil areas in overlapping orthomosaics and through-origin bandwise regressions of the form

$$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f, \qquad \gamma_A = \gamma_a + \gamma_s,$$0

Each target image is calibrated by dividing by the fitted slope, and the references are chained sequentially from one date to the next. In that dataset, calibrated RGB time-series orthomosaics supported an overall accuracy of 90.7% with ensemble (subspace discriminant), compared with 83.0% for the strongest uncalibrated result, which the paper summarizes as an improvement of approximately 7% (Thiruchittampalam et al., 2024).

A more explicitly probabilistic treatment appears in robust relative radiometric normalization for satellite imagery. There the observation model is

$$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f, \qquad \gamma_A = \gamma_a + \gamma_s,$$1

with a latent binary variable distinguishing no-change from change points. The method uses the prior that no-change points should have small residual noise after normalization, whereas change points should have large residual noise. The best-performing variant, HM-RRN-MoG, combines histogram matching with a two-component Gaussian residual model and produces a no-change set that can be evaluated by a no-change set root mean square error. The paper reports that this model is robust against clouds, fogs, and changes, and that it reduces radiometric contrast and NDVI/NDWI differences on the no-change set (Liu et al., 2021).

Exposure behavior can itself be a source of radiometric inconsistency. A multispectral UAV study with a MicaSense RedEdge-3 shows that auto-exposure changes exposure time and gain by band and by scene content, and that this degrades orthomosaic fidelity even after panel and irradiance normalization. Its raw-to-radiance model includes black-level subtraction, gain normalization, vignette correction, and exposure-time normalization,

$$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f, \qquad \gamma_A = \gamma_a + \gamma_s,$$2

followed by panel-based and object-based empirical line calibration. The study finds that fixed exposure orthomosaics were closer to ground truth than auto-exposure orthomosaics, with higher $$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f, \qquad \gamma_A = \gamma_a + \gamma_s,$$3 and lower MAPE, and identifies ideal exposure ranges within which MAPE was $$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f, \qquad \gamma_A = \gamma_a + \gamma_s,$$4 for blue, green, red, and NIR and $$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f, \qquad \gamma_A = \gamma_a + \gamma_s,$$5 for red edge; beyond the upper limit, MAPE increased exponentially. In downstream nitrogen-uptake prediction, fixed exposure gave mostly $$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f, \qquad \gamma_A = \gamma_a + \gamma_s,$$6 with MAPE $$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f, \qquad \gamma_A = \gamma_a + \gamma_s,$$7 to $$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f, \qquad \gamma_A = \gamma_a + \gamma_s,$$8, whereas auto-exposure gave mostly $$\frac{dL_f(\theta,\phi)}{dz} = -\gamma_A L_f + J_f, \qquad \gamma_A = \gamma_a + \gamma_s,$$9 with MAPE $\mathrm{O_2, N_2, O_3, CO_2, H_2O}$0 to $\mathrm{O_2, N_2, O_3, CO_2, H_2O}$1 (Swaminathan et al., 2024).

These workflows share a common structure. Radiometric consistency is not treated as a single gain-offset correction, but as a property that depends on invariant targets, acquisition protocol, exposure stability, and the explicit separation of no-change radiometric variation from real scene change. This suggests that “calibration” and “consistency” are separable: calibration is the procedure, while consistency is the cross-scene condition the procedure is meant to restore.

3. Cross-sensor normalization, reference domains, and mosaics

In some applications, especially mosaicking, radiometric consistency is defined operationally rather than physically. A lunar mosaicking study using Chandrayaan-2 TMC and SELENE imagery defines consistency as the condition in which adjacent tiles exhibit compatible brightness, contrast, and tonal response so that the mosaic is illumination-consistent and free of obvious seam lines. The proposed framework learns a nonlinear mapping from a conventionally mosaicked multi-mission image $\mathrm{O_2, N_2, O_3, CO_2, H_2O}$2 to a normalized image $\mathrm{O_2, N_2, O_3, CO_2, H_2O}$3 aligned with an LROC WAC reference:

$\mathrm{O_2, N_2, O_3, CO_2, H_2O}$4

Its cGAN objective combines an adversarial term with an $\mathrm{O_2, N_2, O_3, CO_2, H_2O}$5 term and reconstructs large mosaics by overlap-aware blending,

$\mathrm{O_2, N_2, O_3, CO_2, H_2O}$6

The best reported checkpoint occurs at epoch 125, with PSNR $\mathrm{O_2, N_2, O_3, CO_2, H_2O}$7 dB and SSIM $\mathrm{O_2, N_2, O_3, CO_2, H_2O}$8, and the paper interprets the method chiefly as seam suppression and tonal harmonization anchored to a radiometrically stable reference domain rather than as strict physical reflectance retrieval (Singh et al., 28 Apr 2026).

A related but distinct problem arises in super-resolution when the supervisory sensor is not the radiometrically preferred one. In multi-image super-resolution from Sentinel-2 to PlanetScope resolution, the paper argues that naive supervised training makes the output “look like” PlanetScope, even though Sentinel-2 Level-2A provides bottom-of-atmosphere reflectance with stronger radiometric calibration and higher radiometric depth. It therefore introduces a radiometric consistency module that constrains the super-resolved image $\mathrm{O_2, N_2, O_3, CO_2, H_2O}$9 to downsample back to the Sentinel-2 reference frame while a separate branch learns a PlanetScope-like appearance. The reconstructed loss described in the paper is

$E_I = E_R + E_A + E_T,$0

The paper shows qualitatively and by histograms that the consistency-constrained outputs preserve Sentinel-2 band distributions rather than drifting toward PlanetScope-like color statistics (Razzak et al., 2021).

A plausible implication is that reference domains do not always serve the same role. In the lunar cGAN, the reference domain defines the desired appearance of the normalized mosaic. In Sentinel-2 super-resolution, the radiometrically trusted source domain remains the target for spectral fidelity, while the high-resolution reference is used mainly for spatial supervision. Radiometric consistency therefore depends not only on a mapping function, but on which sensor is taken to be the authoritative radiometric source.

4. Instrument calibration, channel stability, and mission-level consistency

At the instrument level, radiometric consistency means that detector counts can be mapped to physical units in a stable and cross-channel coherent way. For the New Horizons Multispectral Visible Imaging Camera, this was established with two semi-independent in-flight techniques: stellar calibration for all detectors and a relative Charon-based calibration for the blue, near-infrared, and methane channels scaled from the red channel. The two methods agreed to better than 7%, and the stellar method was adopted for the radiometric keywords delivered to the Planetary Data System because it covers all detectors and does not require a color target in the field of view. The same study also reports gain drift in the near-infrared detector and in one panchromatic framing path, with image-to-image bootstrapping used as a mitigation for affected observations (Howett et al., 2016).

Hinode/EIS presents a different failure mode: degradation of one channel relative to another. A revised in-flight calibration based on line-ratio diagnostics concludes that the long-wavelength channel suffered significant degradation with time, such that by the beginning of 2010 its responsivity was already a factor of two or more below the values measured on the ground. The authors propose a first-order correction under which the main SW/LW ratios become constant to within a relative 20%, the quiet-Sun He II 256 Å radiances become constant over time, and long-standing discrepancies involving strong Fe X, Fe XIII, Fe XIV, Fe XVII, and Fe XXIV lines are largely removed (Zanna, 2012).

Mission-level radiometric consistency can also be framed as a common forward model for signal and noise. ArielRad, developed for the Ariel exoplanet mission, propagates each target stellar spectrum through a shared payload model, computes count rates channel by channel, and then evaluates uncertainty with a physically motivated noise model plus margins for correlated and time-dependent terms. Its total variance per spectral bin is summarized as a combination of photon noise, detector noise, jitter noise, and a payload floor. The paper concludes that Ariel measurements are generally photon-noise dominated, especially in the infrared, which is precisely the regime in which repeatable, comparable radiometric performance is most straightforward to maintain (Mugnai et al., 2020).

An even stricter notion of consistency appears in quantum radiometric calibration. Using squeezed vacuum states and the Heisenberg uncertainty relation, the method infers total efficiency from measured quadrature variances,

$E_I = E_R + E_A + E_T,$1

and then factors out non-detector losses to obtain detector efficiency in situ at the application sideband frequency. For high-efficiency InGaAs photodiodes at 1550 nm, the reported values are

$E_I = E_R + E_A + E_T,$2

at 5 MHz. Here radiometric consistency is not simply responsivity at DC, but frequency-specific fidelity of optical-to-electrical conversion for quantum-noise measurements (Albers et al., 16 Dec 2025).

5. Computer vision, active imaging, and inverse rendering

In computer vision, radiometric consistency often denotes a single physically compatible explanation of appearance across images or directions. “Radiometric Scene Decomposition” formulates this as joint inference of reflectance $E_I = E_R + E_A + E_T,$3, illumination $E_I = E_R + E_A + E_T,$4, and geometry $E_I = E_R + E_A + E_T,$5 from RGB-D observations $E_I = E_R + E_A + E_T,$6,

$E_I = E_R + E_A + E_T,$7

with a rendering model that includes non-Lambertian BRDFs, visibility, shadows, and interreflection. The central consistency condition is that all HDR RGB observations of a scene be explained by one coherent set of scene variables under a physically based rendering equation, rather than by separate per-view photometric adjustments (Lombardi et al., 2016).

A more explicit physical constraint appears in inverse global illumination using a neural radiometric prior. There the learned radiance field $E_I = E_R + E_A + E_T,$8 is penalized by the residual of the rendering equation,

$E_I = E_R + E_A + E_T,$9

and the inverse-rendering objective jointly minimizes photometric error and the norm of this residual. The paper’s interpretation is that image agreement alone is insufficient: the radiance field should also satisfy a transport equilibrium condition so that global illumination is not baked into material or lighting estimates (Hadadan et al., 2023).

Radiometrically Consistent Gaussian Surfels takes the same idea into Gaussian-splatting inverse rendering. It defines a residual between each Gaussian surfel’s learned outgoing radiance and a physically based rendered radiance,

$$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$0

with loss

$$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$1

The point is to supervise unobserved directions that matter for indirect illumination. In ablation, removing $$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$2 drops NVS PSNR from 37.86 to 35.82, albedo PSNR from 31.05 to 30.82, and relight PSNR from 32.09 to 31.69, which the paper interprets as direct evidence that radiometric consistency improves inter-reflection modeling and disentanglement (Han et al., 2 Mar 2026).

Active imaging systems introduce another variant. In camera-display communication, the relevant transfer is not scene reflectance to camera response but display emittance to camera response. The paper models the inverse radiometric function as a fourth-order polynomial,

$$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$3

then derives a physics-based feature map for simultaneous radiometric calibration and message recovery. Under the difficult case of 45° viewing angle and additive difference +3, naive thresholding averages 56.17% accuracy, whereas optimal online radiometric calibration reaches 98.11% across nine commercial camera-display pairs (Yuan et al., 2015).

6. Thermal, medical, and chronometric extensions

Thermal benchmark construction has made radiometric consistency an annotation principle rather than an image-correction procedure. In FlameVQA, the radiometric thermal TIFF—not the colorized thermal JPEG—is the authoritative thermal signal throughout dataset construction, supervision, and quality control. The TIFF is described as a single-band raster with per-pixel temperature values, and hotspot logic is based on a fixed physical threshold,

$$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$4

with $$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$5 in $$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$6. Labels for temperature-critical questions are verified or overridden by deterministic TIFF-based rules, and the benchmark treats this as radiometric thermal supervision rather than palette-based interpretation (Habibpour et al., 25 Jun 2026).

WildFireVQA develops the same principle at larger scale. Each sample includes an RGB image, a color-mapped thermal visualization, and a radiometric thermal TIFF, and the benchmark states that every pixel in the TIFF encodes an absolute temperature in degrees Celsius. Its deterministic hotspot mask is

$$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$7

with further physical normalization by ground sampling distance,

$$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$8

to make hotspot area and spacing comparable across UAV altitudes. Retrieval-augmented settings append radiometric statistics such as minimum, maximum, standard deviation, and percentages of pixels above $$\rho_\lambda = \frac{E_R(\lambda)}{E_I(\lambda)} \times 100.$$9 and $$BRDF = \frac{dL_r(\theta_r,\phi_r;\theta_i,\phi_i)}{dE_i},$$0, making explicit the difference between visually encoded heatmaps and quantitative temperature fields (Habibpour et al., 22 Apr 2026).

Outside optical and thermal imaging, the same concern appears as reproducibility of quantitative measurements under acquisition change. A CT study reconstructs raw data from 23 lung nodules under 320 combinations of dose, kernel, and slice thickness and defines “compatible pairs” as pairs of reconstruction conditions that produce no significant change in a feature. Its central result is a trade-off: as thickness increases, volumetric reproducibility decreases, while reproducibility of histogram- and texture-based features improves. The authors conclude that balanced standardization of acquisition parameters is required if volumetric and radiomic results are to be concomitantly reproducible across studies (Erdal et al., 2019).

In meteoritic chronology, radiometric consistency means concordance among independent clocks. By fitting the calibration quantities that link Al–Mg, Mn–Cr, Hf–W, and Pb–Pb ages to a common Solar System zero point, the study finds that 37 of 38 formation times across 14 achondrites can be made concordant with

$$BRDF = \frac{dL_r(\theta_r,\phi_r;\theta_i,\phi_i)}{dE_i},$$1

using

$$BRDF = \frac{dL_r(\theta_r,\phi_r;\theta_i,\phi_i)}{dE_i},$$2

The paper interprets this as strong support for homogeneity of $$BRDF = \frac{dL_r(\theta_r,\phi_r;\theta_i,\phi_i)}{dE_i},$$3, $$BRDF = \frac{dL_r(\theta_r,\phi_r;\theta_i,\phi_i)}{dE_i},$$4, and $$BRDF = \frac{dL_r(\theta_r,\phi_r;\theta_i,\phi_i)}{dE_i},$$5 in the solar nebula (Desch et al., 2022).

Across these domains, a common misconception is that radiometric consistency is synonymous with a one-time gain correction. The literature instead treats it as a broader invariance condition: invariance to atmosphere and geometry in remote sensing, to exposure and gain selection in UAV imaging, to cross-sensor domain shift in mosaics and super-resolution, to detector drift and channel imbalance in instrumentation, to unobserved transport directions in inverse rendering, and to protocol variation in quantitative imaging. A plausible implication is that consistency is best understood not as a single algorithmic step, but as an end state in which the quantities used for scientific inference remain physically interpretable under the transformations that matter for a given measurement regime.