Organizations: College of Electrical Engineering, Nanjing Institute of Technology, Nanjing, Jiangsu 210000 · College of Communication Engineering, Nanjing Institute of Technology, Nanjing, Jiangsu 210000 · College of Resource and Environmental, Anhui University, Hefei, Anhui 230601
Defense Meteorological Satellite Program (DMSP-OLS) and Suomi National Polar-orbiting Partnership (SNPP-VIIRS) nighttime light (NTL) data are vital for monitoring urbanization, yet sensor incompatibilities hinder long-term analysis. This study proposes a cross-sensor calibration method using Contrastive Unpaired Translation (CUT) network to transform DMSP data into VIIRS-like format, correcting DMSP defects. The method employs multilayer patch-wise contrastive learning to maximize mutual information between corresponding patches, preserving content consistency while learning cross-domain similarity. Utilizing 2012-2013 overlapping data for training, the network processes 1992-2013 DMSP imagery to generate enhanced VIIRS-style raster data. Validation results demonstrate that generated VIIRS-like data exhibits high consistency with actual VIIRS observations (R-squared greater than 0.87) and socioeconomic indicators. This approach effectively resolves cross-sensor data fusion issues and calibrates DMSP defects, providing reliable attempt for extended NTL time-series.
Figures & tables
Figure 1: NTL visualization and cumulative analysis of longitude and latitude
Parameter
DMSP-OLS
Suomi NPP VIIRS
Operator
U.S. Air Force
NASA-NOAA (JPSS)
Update period
Annual
monthly
Orbit
Polar Sun-Synchronous (850 km, 98.8∘ )
Polar Sun-Synchronous (827 km, 98.7∘ )
Revisit Time
102 minutes
102 minutes
Swath Width
3,000 km
3,000 km
Local Overpass Time
∼ 19:30 (Descending, pre-2013)
∼ 01:30 (Descending)
Table 1: Technical comparison between DMSP-OLS and Suomi NPP VIIRS sensors.
Figure 2: NTL data from DMSP-OLS and NPP-VIIRS
Figure 3: Schematic of the preprocessing workflow. The pipeline consists of resolution alignment, patch extraction with spatial filtering, radiometric distribution analysis, and format conversion for network training.
Property
Original DMSP
Upsampled DMSP
Pixel size (°)
0.0083333333
0.0041666667
Width (pixels)
43,201
86,402
Height (pixels)
16,801
33,602
Total pixels
725.8 M
2,903.2 M
Data type
Float32
Float32
Table 2: Resampling parameters and output verification for DMSP-OLS upsampling.
Figure 4: Radiometric distribution analysis. Top: histograms with log1p transformation to visualize long-tail behavior. Bottom: cumulative distribution functions. Vertical dashed lines indicate the 99.9th percentile thresholds used for outlier clipping.
Split
Number of Patches
Geographic Coverage
Purpose
Training
57,316
Global (stratified)
Model optimization
Validation
14,329
Global (stratified)
Hyperparameter tuning
Test
14,329
Held-out regions
Final evaluation
Total
12,417
Global land areas
Table 3: Final dataset composition after preprocessing and spatial partitioning.
Property
DMSP-OLS
VIIRS DNB
Resolution (after alignment)
500 m
500 m
Image dimensions
86,402 × 33,602
86,401 × 33,601
Raw value range
[0.00, 63.00]
[ − 1.50, 85,588.98]
log1p 99.9th percentile
4.1589
4.3643
Quantization
6-bit (saturated)
14-bit (linear)
Table 4: Radiometric and spatial characteristics of DMSP-OLS and VIIRS DNB data after preprocessing.
Figure 5: Overall network pipeline. The generator G translates DMSP to VIIRS-like imagery, while the discriminator distinguishes real from fake samples. Multilayer patchwise contrastive loss operates on intermediate features to ensure spatial consistency without requiring cycle reconstruction.
Figure 6: Patchwise contrastive learning mechanism. Features are extracted from both domains at multiple network depths, projected through MLP heads into a 256-dimensional embedding space, and compared via softmax cross-entropy. Positive pairs (blue) come from the same spatial location; negative pairs (yellow) come from different locations.
Parameter
Value
Optimizer
Adam ( β1=0.5,β2=0.999 )
Initial learning rate
2×10−4
LR schedule
Linear decay after epoch 200
Total epochs
400
Batch size
4
Patch size
256×256
Table 5: Training hyperparameters for the CUT-based calibration model.
Figure 7: Training loss convergence. Top: PatchNCE loss decreases from 4.5 to 0.5. Bottom: GAN loss stabilizes around 0.4–0.8 after epoch 500.
Figure 8: Density scatter plot of generated vs. real VIIRS radiance values. Color indicates point density (yellow = high, purple = low). The white line represents perfect 1:1 correspondence.
Radiance Range
Pixel Count
MAE
RMSE
R2
0–20 (Background)
1,245,832
2.31
3.87
0.91
20–40 (Rural)
423,156
4.12
6.23
0.88
40–60 (Suburban)
187,429
6.45
9.18
0.85
60–80 (Urban)
64,821
9.87
14.52
0.79
> 80 (Urban Core)
28,394
15.23
21.67
0.68
Overall
1,949,632
5.64
8.91
0.87
Table 6: Error metrics stratified by radiance intensity level.
Method
R2
Spearman ρ
SSIM
Training Time
Linear Regression
0.72
0.74
0.61
N/A
Histogram Matching
0.68
0.71
0.58
N/A
CycleGAN
0.81
0.83
0.74
95 hours
CUT (Ours)
0.87
0.91
0.79
80 hours
Table 7: Comparison with baseline calibration methods. Best values in bold.
Figure 9: Successful calibration examples from 2012 test set. Left column of each pair: original DMSP. Right column: generated VIIRS-like output. Rows show diverse settlement patterns from coastal cities to inland towns.
Figure 10: Failure mode examples. Left: original DMSP. Right: generated output. Artifacts include checkerboard patterns in saturated cores (top row) and halo effects around bright spots (bottom row).
Metric
Value
R2 (radiance correlation)
0.87
RMSE (normalized)
0.114
SSIM (structural similarity)
0.79
Successful cases (visual)
∼ 75%
Failure cases (artifacts)
∼ 8%
Table 8: Quantitative evaluation metrics on held-out test set (n = 1,862 patches).
School of Artificial Intelligence and Computer Science, Jiangnan University · Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey
MoE Key Lab of Artificial Intelligence, Institute of AI, School of Computer Science, Shanghai Jiao Tong University, Shanghai, China · Tsinghua University, Beijing, China · Institute of Modern Optics, Nankai University, Tianjin, China