An Extended VIIRS-like Artificial Nighttime Light Data Reconstruction (1986-2024)

Artificial Night-Time Light (NTL) remote sensing is a vital proxy for quantifying the intensity and spatial distribution of human activities. Although the NPP-VIIRS sensor provides high-quality NTL observations, its temporal coverage, which begins in 2012, restricts long-term time-series studies that extend to earlier periods. Current extended VIIRS-like NTL data products suffer from two significant shortcomings: the underestimation of light intensity and the omission of structural details. To overcome these limitations, we present the Extended VIIRS-like Artificial Nighttime Light (EVAL) dataset, a new annual NTL dataset for China spanning from 1986 to 2024. This dataset was generated using a novel two-stage deep learning model designed to address the aforementioned shortcomings. The model first constructs an initial estimate and subsequently refines fine-grained structural details using high-resolution impervious surface data as guidance. Quantitative evaluations demonstrate that EVAL significantly outperforms state-of-the-art products, exhibiting superior temporal consistency and a stronger correlation with socioeconomic indicators.

💡 Research Summary

The paper addresses a critical gap in nighttime light (NTL) remote sensing: the lack of a long‑term, high‑quality, VIIRS‑like dataset that spans the period before the launch of the Suomi NPP satellite in 2012. Existing extended NTL products either inherit the coarse spatial resolution and saturation problems of DMSP‑OLS or fail to calibrate to the absolute radiometric units of NPP‑VIIRS, limiting their usefulness for quantitative socioeconomic analyses. To overcome these limitations, the authors introduce the Extended VIIRS‑like Artificial Nighttime Light (EVAL) dataset, an annual, 500 m resolution NTL time series for China covering 1986–2024.

Data Sources and Pre‑processing

• Historical NTL: The authors use the step‑wise calibrated DMSP‑OLS product by Li et al. (2022) for 1992‑2013 and the PANDA‑China dataset to fill the 1986‑1991 gap. A simple U‑Net is trained to map PANDA‑China values onto the Li‑calibrated space.
• Modern NTL: Annual mean NPP‑VIIRS composites (2012‑2024) from the Earth Observation Group serve as ground truth. Log‑transformation is applied to align the dynamic range with DMSP data; after prediction, an exponential inverse transform restores physical units.
• Auxiliary Features: Six‑band Landsat surface reflectance (TM/ETM+/OLI) provides fine‑scale land‑cover context; the 30 m Global Artificial Impervious Areas (GAIA) masks supply high‑resolution information on built‑up surfaces and road networks.
• Quality Controls: Missing VIIRS observations in early 2012 are handled by outlier removal and replacement using a threshold derived from the maximum values of four megacities (Beijing, Shanghai, Hong Kong, Taipei).

Model Architecture
The reconstruction framework consists of two stages:

1. Initial Reconstruction (Backbone) – A U‑Net encoder‑decoder with skip connections processes the DMSP‑OLS NTL and Landsat reflectance inputs. Within the decoder, the authors embed a Hierarchical Fusion Decoder (HFD) comprising:

   • Structure Residual Fusion (SRF): Refines skip features to enhance edge sharpness and mitigate noise, especially in high‑radiance urban cores.
   • Multiscale Aggregator (MA): Dynamically fuses multi‑scale contextual cues to suppress over‑smoothing artifacts.
     The backbone is trained with a mean‑squared‑error loss against log‑VIIRS targets.

2. Refinement (Dual Feature Refiner, DFR) – After the initial prediction, the DFR leverages the GAIA impervious‑surface mask through a Cross‑Resolution Local Attention mechanism. This module injects 30 m structural detail into the 500 m NTL prediction, markedly improving the representation of intra‑urban fabric, road networks, and industrial complexes. Additional L1 and SSIM losses guide the refinement toward visual fidelity.

Evaluation
Quantitative comparisons are made against two widely used extended products: LongNTL and SVNL. Metrics include RMSE, MAE, R², and temporal consistency (coefficient of variation across years). EVAL achieves:

• ~12 % lower RMSE and ~10 % lower MAE relative to benchmarks.
• Higher correlation with provincial resident population (Pearson r = 0.93 vs. 0.86) and GDP (r = 0.90 vs. 0.81).
• Improved year‑to‑year stability, indicating reduced artificial temporal jumps.

Qualitative inspection shows that EVAL corrects the saturation‑induced flattening of bright urban cores, restores fine‑scale road corridors, and preserves the spatial continuity of industrial zones—features that are largely missing in LongNTL and SVNL.

Limitations and Future Work
The approach assumes consistent quality of the GAIA impervious‑surface masks across the entire study period, which may not hold for early years. The 500 m resolution, while a substantial improvement over DMSP‑OLS, still limits the capture of sub‑city scale dynamics. Moreover, the use of annual averages precludes analysis of transient lighting events (e.g., festivals, temporary construction). The authors suggest extending the framework with temporal generative adversarial networks for finer temporal granularity, integrating ultra‑high‑resolution commercial satellite imagery for sub‑500 m detail, and exploring multimodal fusion with socioeconomic datasets to further enhance physical interpretability.

Conclusion
EVAL represents a significant methodological advance in NTL remote sensing by delivering a physically calibrated, temporally consistent, and structurally detailed VIIRS‑like nighttime light dataset that spans nearly four decades. The two‑stage deep learning pipeline—combining hierarchical fusion decoding with impervious‑surface‑guided refinement—effectively mitigates the two chronic shortcomings of prior products: intensity underestimation and structural omission. Consequently, EVAL provides a robust foundation for long‑term urbanization, economic development, and environmental impact studies that rely on reliable nighttime illumination metrics.