CNN-Based Spectral Super-Resolution of Panchromatic Night-Time Light Imagery: City-Size-Associated Neighborhood...

Area of Science:

• Earth Observation
• Remote Sensing
• Artificial Intelligence

Background:

• Satellite-derived panchromatic night-time light (NTL) data is globally available.
• Spectral NTL data offers richer analytical information but is often localized or commercial.
• Previous work explored machine learning for panchromatic to spectral NTL conversion.

Purpose of the Study:

• To investigate the neighborhood effect in spectral NTL prediction using convolutional neural networks (CNNs).
• To determine the optimal input image size for CNN models based on metropolitan area extent and light color.
• To compare CNN performance against other machine learning techniques for NTL spectral conversion.

Main Methods:

• Utilized convolutional neural networks (CNNs) to analyze the spatial context of NTL data.
• Explored the relationship between input image size, metropolitan area size, and spectral NTL prediction accuracy.
• Compared CNN performance with linear regression, kernel regression, random forest, and elastic map models.

Main Results:

• The neighborhood effect is significant and scales with the geographical extent of metropolitan areas.
• Optimal input image size for CNNs is smaller for smaller cities and varies by color for larger cities (higher for red, lower for blue).
• CNNs achieved comparable Pearson's correlation and superior Weighted Mean Squared Error (WMSE) compared to other methods, particularly on testing datasets.

Conclusions:

• CNNs effectively capture the neighborhood effect for spectral NTL estimation.
• The optimal input size for CNNs in spectral NTL analysis is dependent on city size and light color.
• CNNs offer a promising approach for enhancing the analysis of global spectral NTL data.