Extended VIIRS-like Artificial Nighttime Light (EVAL)

• EVAL is a methodological family that extends, calibrates, and reconstructs VIIRS nighttime imagery using deep learning and cross-domain translation techniques.
• It employs conditional GANs, two-stage deep architectures, and unpaired calibration methods to harmonize multisensor data from Landsat, DMSP-OLS, and ISS imagery.
• Validation metrics such as R², RMSE, and PSNR demonstrate that EVAL improves temporal continuity and spatial accuracy for urban and environmental analyses.

to=arxiv_search ็ตทรู 大奖彩票站 天天中彩票中大奖 大发快三彩票json {"query":"Extended VIIRS-like Artificial Nighttime Light EVAL nighttime light VIIRS-like arXiv", "max_results": 10} to=arxiv_search ՞նչ 重庆时时彩彩ೆjson {"query":"nighttime light VIIRS-like reconstruction arXiv DMSP VIIRS EVAL", "max_results": 10} to=arxiv_search 开号网址 彩神争霸可以json {"query":"Translating multispectral imagery to nighttime imagery via conditional generative adversarial networks arXiv", "max_results": 5} Extended VIIRS-like Artificial Nighttime Light (EVAL) denotes a set of related nighttime-remote-sensing constructs that seek to extend, reconstruct, calibrate, or enrich VIIRS-style observations beyond the native Suomi-NPP VIIRS record. In the cited literature, the term covers several distinct but connected objectives: translating Landsat 8 multispectral imagery to VIIRS-style nighttime radiance with a conditional GAN; proposing a roadmap for synthetic “VIIRS-style” annual composites from historical Landsat archives; reconstructing a China-wide 1986–2024 product with a two-stage deep architecture; calibrating DMSP-OLS into VIIRS-like format with unpaired cross-domain translation; generating visible multispectral nighttime products from calibrated ISS DSLR imagery; and defining a composite metric that combines orbital and human-vision sensitivities (Huang et al., 2019, Tian et al., 1 Aug 2025, Tong et al., 17 Mar 2026, Miguel et al., 2021, Bará et al., 2024).

1. Conceptual scope and motivation

Nighttime satellite imagery has been applied in a wide range of fields, but its further application is hindered by limited understanding of how observed light intensity is formed and by uncertainty over whether it can be simulated (Huang et al., 2019). A central motivation for EVAL is temporal extension: the NPP-VIIRS sensor provides high-quality NTL observations, yet its temporal coverage begins in 2012, which restricts long-term time-series studies that extend to earlier periods (Tian et al., 1 Aug 2025). A second motivation is sensor incompatibility. DMSP-OLS and SNPP-VIIRS nighttime-light data are both widely used for monitoring urbanization, but discontinuities in calibration, spatial resolution, and saturation behavior hinder long-term fusion and interannual analysis (Singh et al., 2023, Tong et al., 17 Mar 2026).

A third motivation is spectral incompleteness. VIIRS/DNB is panchromatic, and DMSP/OLS and SNPP/VIIRS-DNB are panchromatic and multispectral in the infrared but not at visible wavelengths, whereas ISS DSLR imagery can provide visible multispectral data across the world, albeit with substantial processing requirements (Miguel et al., 2021). Related work on colorizing panchromatic VIIRS/DNB into RGB proxies makes the same point in a different form: panchromatic radiance masks spectral signatures associated with lighting technologies and land-use types, limiting urban, health, and ecological analyses (Rybnikova et al., 2020).

The literature therefore uses EVAL in more than one technical sense. Sometimes it denotes a raster reconstruction of VIIRS-like radiance from older sensors; sometimes it denotes a multispectral or RGB augmentation of panchromatic NTL; sometimes it denotes a physically calibrated ground-illumination product; and sometimes it denotes a composite metric designed to track both orbital radiance and human-vision radiance as lamp spectra evolve (Buhler et al., 3 Oct 2025, Bará et al., 2024). This suggests that EVAL is best understood as a methodological family rather than a single universally standardized product.

2. Conditional translation from multispectral imagery to VIIRS-style radiance

A foundational precursor to later EVAL work is Huang et al.’s use of a modified pix2pix conditional GAN to translate Landsat 8 multispectral imagery to nighttime imagery (Huang et al., 2019). The training data were gridded Landsat 8 and VIIRS image pairs over CONUS for 2016. Landsat 8 Collection-1 Tier 1 surface-reflectance Bands 4, 3, 2, and 5 were obtained via Google Earth Engine; each scene was atmospherically corrected, orthorectified, cloud-masked, median-composited per pixel, and resampled to 100 m. The VIIRS NPP-VJ1 DNB annual stable-light composite for 2016 was downloaded from NOAA/NCEI; stray light, lightning, lunar and cloud contamination were removed, transient lights and non-lights were screened, and radiance values were capped at 300 nW cm$^{-2}$ sr$^{-1}$. Both datasets were reprojected to a common WGS84 100 m grid, tiled into $0.625^\circ \times 0.625^\circ$ grids, and converted into aligned $256 \times 256$ patches (Huang et al., 2019).

The network used a U-Net generator and a $70 \times 70$ PatchGAN discriminator. The generator comprised seven convolutional encoder blocks, one 512-filter bottleneck convolution, and seven deconvolutional decoder blocks with skip-connections. Input channels were configured as RGB only, RGB + IR, or RGB + IR + SM, where the “social media” channel was derived from geotagged tweets from July to December 2016, aggregated into $\sim 450$ m cells, log-transformed, low-pass filtered with a $3 \times 3$ kernel, resampled to 100 m, and normalized to $[-1,1]$. The discriminator used five convolutional blocks with LeakyReLU $(\alpha = 0.2)$ and a final Sigmoid patch-probability map. The objective combined the adversarial loss and an $L_1$ reconstruction penalty with $^{-1}$0:

$^{-1}$1

$^{-1}$2

$^{-1}$3

Training used 1000 screened CONUS grids, selected by total VIIRS radiance $^{-1}$4 nW cm$^{-1}$5 sr$^{-1}$6, with 800 for training and 200 for validation. Hyperparameters followed the original pix2pix setup: batch size 1, Adam with $^{-1}$7, $^{-1}$8, $^{-1}$9, and 200 epochs, with the first 100 at constant learning rate and the next 100 linearly decayed to zero (Huang et al., 2019).

The quantitative results established the practical viability of multispectral-to-nighttime translation.

Scenario	Deu / Dma	$0.625^\circ \times 0.625^\circ$0
RGB	31.431 / 41.947	0.492
RGB + IR	24.872 / 31.826	0.647
RGB + IR + SM	19.214 / 23.234	0.821

Visual comparison showed that RGB-only input failed to reproduce urban brightness, RGB + IR captured built-up areas but missed intensity cores, and RGB + IR + SM most closely matched VIIRS ground truth, including shadows, urban cores, and noise suppression (Huang et al., 2019). In the EVAL context, Huang et al. explicitly proposed training on the overlapping Landsat 8/VIIRS era and applying the generator to historical Landsat 4/5/7 surface reflectance since 1984 to create synthetic “VIIRS-style” annual composites back to 1984. They also identified multi-sensor calibration, light desaturation, and temporal gap filling as downstream uses (Huang et al., 2019).

3. Reconstruction frameworks for long time series

Subsequent EVAL-oriented work moved from proof-of-concept translation to systematic time-series reconstruction. In the China EVAL product, the reconstruction task is decomposed into a two-stage process: a construction stage and a refinement stage (Tian et al., 1 Aug 2025). The construction stage uses a U-Net encoder-decoder with a ResNet-50 encoder, a Structure Residual Fusion module, and a Multi-Scale Aggregator with dilated convolutions at rates $0.625^\circ \times 0.625^\circ$1. It is optimized with mean squared error on log-transformed VIIRS radiance,

$0.625^\circ \times 0.625^\circ$2

while the refinement stage uses a Dual Feature Refiner guided by high-resolution impervious surface masks and Cross-Resolution Local Attention, producing a residual correction so that

$0.625^\circ \times 0.625^\circ$3

The refinement loss is $0.625^\circ \times 0.625^\circ$4 on log-space prediction. Inputs include Harmonized Global NTL DMSP-OLS, PANDA-China NTL, NPP-VIIRS annual composites, Landsat surface reflectance, GAIA annual impervious masks, and socioeconomic data. Preprocessing includes annual 10th-percentile Landsat compositing, OLI to TM/ETM+ harmonization via Roy et al. (2016) coefficients, resampling to 250 m, log-transforming VIIRS targets during training, per-band min–max normalization to $0.625^\circ \times 0.625^\circ$5, and sampling 20,000 patches of $0.625^\circ \times 0.625^\circ$6 pixels with an 80/10/10 train/val/test split by province (Tian et al., 1 Aug 2025).

A different paradigm appears in the CUT-based unpaired cross-domain calibration of DMSP to VIIRS (Tong et al., 17 Mar 2026). Here the objective is not paired Landsat-to-VIIRS translation but DMSP-to-VIIRS transformation using 2012–2013 overlap. DMSP annual composites from 1992–2013 are upsampled from 30″ to 15″ by bilinear interpolation, log-transformed, clipped at the 0.1%–99.9% percentiles, normalized to uint8, tiled into non-overlapping $0.625^\circ \times 0.625^\circ$7 patches, and filtered to exclude patches with less than 30% land, centered at $0.625^\circ \times 0.625^\circ$8, or radiometrically dark or uniform. The generator is a ResNet-style image-to-image network with 9 residual blocks; the discriminator is a $0.625^\circ \times 0.625^\circ$9 PatchGAN; and the defining component is PatchNCE contrastive learning over generator features extracted from layers $256 \times 256$0, with $256 \times 256$1 spatial samples per layer and temperature $256 \times 256$2. The total objective is

$256 \times 256$3

with $256 \times 256$4 and $256 \times 256$5 (Tong et al., 17 Mar 2026).

Singh et al.’s NDUI+ work addresses inter-sensor normalization from a different direction: learning a nonlinear mapping from raw VIIRS DNB radiance to DMSP-OLS–style radiance using 2012 overlap, then fusing the calibrated VIIRS output with Landsat 7 NDVI into a 30 m annual urban index from 1999 to the present (Singh et al., 2023). The globe is divided into $256 \times 256$6 blocks, a separate model is trained for each block, and the Swin Transformer is chosen after benchmarking 17 super-resolution and sensor-fusion networks in the UT-CDS suite. The loss combines MSE with a small gradient penalty for smoothness, with $256 \times 256$7. Although NDUI+ is not itself a radiance-only EVAL product, it exemplifies the same core design pattern: one-shot overlap transfer, regional tuning, and multi-source harmonization of DMSP and VIIRS into a longer, higher-resolution annual series (Singh et al., 2023).

4. Validation, quantitative performance, and temporal consistency

Quantitative evaluation is central to EVAL because the target variables differ across formulations: radiance, RGB proxies, urban indices, skyglow, and ground illuminance. In the China EVAL dataset, pixel-scale evaluation on the 2012 test set in log-space compared EVAL with LongNTL and SVNL, using $256 \times 256$8, RMSE, PSNR, and UIQI (Tian et al., 1 Aug 2025).

Product	Pixel-level results (2012)	Aggregate or correlation result
LongNTL	$256 \times 256$9, RMSE $70 \times 70$0, PSNR $70 \times 70$1, UIQI $70 \times 70$2	GDP corr. $70 \times 70$3, Population corr. $70 \times 70$4
SVNL	$70 \times 70$5, RMSE $70 \times 70$6, PSNR $70 \times 70$7, UIQI $70 \times 70$8	GDP corr. $70 \times 70$9, Population corr. $\sim 450$0
EVAL	$\sim 450$1, RMSE $\sim 450$2, PSNR $\sim 450$3, UIQI $\sim 450$4	GDP corr. $\sim 450$5, Population corr. $\sim 450$6

The same work reported that EVAL boosts $\sim 450$7 from 0.686 to 0.809 and lowers RMSE from 1.278 to 0.997 relative to SVNL; at the city aggregate scale across 2,891 counties in China for 2012, EVAL reached $\sim 450$8 with RMSE $\sim 450$9 W/cm$3 \times 3$0sr, and trend comparison showed that EVAL avoided the jump seen in other VIIRS-like products at the 2012 transition (Tian et al., 1 Aug 2025).

The CUT-based DMSP-to-VIIRS calibration also reported strong held-out consistency. On the 15% test patches, Pearson $3 \times 3$1, Spearman $3 \times 3$2, $3 \times 3$3, concordance correlation coefficient $3 \times 3$4, RMSE $3 \times 3$5, MAE $3 \times 3$6, and SSIM $3 \times 3$7; relative to linear regression, histogram matching, and CycleGAN, CUT had the highest reported $3 \times 3$8 and Spearman $3 \times 3$9, and lower training time than CycleGAN. Error stratification showed high accuracy with $[-1,1]$0 in background and rural areas, degrading to $[-1,1]$1 in highly saturated urban cores (Tong et al., 17 Mar 2026).

For NDUI+, model selection was based on PSNR and SSIM on held-out 2012 patches, with the Swin Transformer yielding $[-1,1]$2 versus 0.85 for the next-best model. Independent validation over Austin in 2020 showed $[-1,1]$3 and MAE $[-1,1]$4 in normalized units against Sentinel-2 Dynamic World built-up probability, while comparison with Ghosh et al. (2021) over Las Vegas indicated that the new dataset recovered the post-2016 radiance rebound (Singh et al., 2023). Taken together, these studies indicate that “VIIRS-like” quality is typically assessed through some combination of pixelwise fidelity, structural similarity, aggregate temporal continuity, and correlation with socioeconomic or urban-form variables.

5. Spectral, photometric, and illuminance-oriented extensions

One branch of EVAL research departs from radiance reconstruction and instead extends VIIRS-like products into the visible spectral domain. The calibration of DSLR images from the International Space Station is a multi-stage process: Decodification, Linearity correction, Flat field/Vignetting, Spectral characterization of the channels, Astrometric calibration/georeferencing, Photometric calibration (stars)/Radiometric correction, and Transmittance correction (Miguel et al., 2021). RAW NEF files are converted with DCRAW into 16-bit linear counts for the four Bayer subchannels; high-DN nonlinearity is corrected with an empirical function $[-1,1]$5; flat-field gains are derived from uniform lambertian targets; spectral response functions are linked to Johnson photometry through synthesized DSLR magnitudes; georeferencing uses Astrometry.net followed by ground control points; and radiometric scaling is derived from star-field regression. The worked Madrid example reported a final green-band map ranging from 0 to 300 nW·sr$[-1,1]$6·cm$[-1,1]$7, per-pixel uncertainty of approximately 5% from star-fit residuals and 2–3% from the atmospheric model, and georeferencing RMSE of approximately 4 pixels, corresponding to 0.4 km at nadir (Miguel et al., 2021). The same workflow proposes storing each band as georeferenced 16-bit GeoTIFF or NetCDF, followed by mosaicking and temporal compositing into a time series compatible with VIIRS DNB at 0.5 km resolution (Miguel et al., 2021).

A second spectral extension is panchromatic-to-RGB colorization of VIIRS/DNB (Rybnikova et al., 2020). In that framework, each pixel is represented by a predictor vector consisting of panchromatic radiance, two neighborhood-difference terms, and two HBASE built-up statistics, and separate models are trained for $[-1,1]$8, $[-1,1]$9, and $(\alpha = 0.2)$0. The candidate methods are multiple linear regression, Nadaraya–Watson kernel regression, random forest regression with $(\alpha = 0.2)$1, and the elastic-map manifold method on a $(\alpha = 0.2)$2 grid with bending penalty tuned over $(\alpha = 0.2)$3. Training data came from eight metropolitan areas—Atlanta, Beijing, Haifa, Khabarovsk, London, Naples, Nashville, and Tianjin—using georeferenced ISS photographs and leave-one-city-out evaluation. Random forest and kernel regression achieved the highest training-set correlations and lowest WMSE, linear regression generalized best on unseen cities with test Pearson $(\alpha = 0.2)$4 in the range 0.75–0.85, and elastic-map models had the highest consistency between train and test metrics (Rybnikova et al., 2020). In EVAL terms, this line of work treats the VIIRS-like product as a spectrally enriched proxy rather than a direct replacement for DNB radiance.

A third extension defines EVAL as a composite metric that explicitly mixes orbital and human-vision sensitivities as lighting technology changes (Bará et al., 2024). Using lamp-type spectral libraries $(\alpha = 0.2)$5, total lumen flux per unit area $(\alpha = 0.2)$6, lamp fractions $(\alpha = 0.2)$7, the VIIRS DNB response, and human luminous-efficiency functions, the proposed metric is

$(\alpha = 0.2)$8

Its rate of change inherits both the growth of total lumens and the evolution of lamp mixture. In the VIIRS-vs-Globe-at-Night application, the paper argues that the reported divergence between approximately 2.2%/yr in VIIRS-DNB radiance and approximately 9.6%/yr in GAN-derived artificial radiance could be explained by spectral changes alone under some adaptation conditions, whereas photopic reconciliation would require GAN-specific light sources not captured by VIIRS-DNB (Bará et al., 2024). Here EVAL is not a raster reconstruction but a spectrally corrected time-series functional.

A fourth extension is Otus 3, which transforms VIIRS-based radiance into physically interpretable maps of ground illuminance and skyglow (Buhler et al., 3 Oct 2025). Using Black Marble monthly and yearly composites, angular-dependence products, halo removal, and calibration against 139 Ninox sky-brightness sites across metropolitan France, the model infers ground luminous emittance, direct artificial illuminance, and artificial skyglow illuminance under clear-sky and low-cloud scenarios. The key inversion uses a Garstang-type emission model,

$(\alpha = 0.2)$9

from which direct artificial illuminance follows as $L_1$0, while diffuse skyglow illuminance is obtained by convolving the GLE map with scattering kernels. The resulting France maps showed a 23% reduction in middle-of-night artificial ground illuminance between 2013–2018 and 2019–2024 (Buhler et al., 3 Oct 2025). In this usage, EVAL denotes a calibrated photometric product, not merely a VIIRS-like radiance field.

6. Limitations, misconceptions, and open directions

A common misconception is to treat EVAL as a single standardized dataset. The literature instead assigns the label to several distinct constructs: radiance reconstruction from Landsat or DMSP, visible multispectral calibration from ISS imagery, RGB colorization of panchromatic VIIRS, urban-index fusion products, illuminance maps, and a spectral composite metric (Huang et al., 2019, Miguel et al., 2021, Rybnikova et al., 2020, Buhler et al., 3 Oct 2025, Bará et al., 2024). A plausible implication is that comparisons among EVAL studies require attention to the target variable itself: top-of-atmosphere radiance, log-transformed VIIRS proxy, RGB channel intensity, NDUI scalar, zenith luminance, or ground illuminance.

Method-specific limitations are explicit. Huang et al.’s cGAN was trained on CONUS, so global application requires regional retraining or domain adaptation; it used annual composites only; and the social-media auxiliary channel may under-represent certain areas because of Twitter skew (Huang et al., 2019). In NDUI+, there is no extension before 1999, the output is a single scalar urban–vegetation index rather than subclass labels, the VIIRS-to-DMSP mapping is assumed stationary after 2012, and the 2015 non-urban mask reused for earlier years may over-mask rapidly urbanizing areas (Singh et al., 2023). In the CUT-based calibration, performance degrades in highly saturated urban cores, despite strong global metrics (Tong et al., 17 Mar 2026). In the France illuminance framework, direct illuminance uncertainty is approximately 49%, the minimum detectable ADI is approximately 25 mlux, and single-value aerosol and emission-law classes ignore local heterogeneity (Buhler et al., 3 Oct 2025).

Open directions are also stated directly in the source material. The Landsat-to-VIIRS roadmap proposes alternative auxiliary sources such as OpenStreetMap and mobile-phone data, as well as recurrent cGANs and atmospheric emissions data (Huang et al., 2019). The ISS calibration workflow points toward multispectral, epoch-tagged EVAL products extending visible artificial-light records back two decades and complementing panchromatic VIIRS DNB (Miguel et al., 2021). The Otus 3 study recommends integrating SDGSAT-1 higher-resolution multispectral nighttime data, expanding Ninox measurements to ground lux, and dynamically ingesting local AOD and cloud statistics (Buhler et al., 3 Oct 2025). Across these variants, the unifying research program is not a single algorithmic template but a persistent attempt to reconcile temporal depth, sensor continuity, spectral fidelity, and physical interpretability in nighttime-light remote sensing.