Nighttime Light Image Reconstruction Using Semantic Constraints

The patent introduces light semantics information (impervious surface data and road network data) as an intermediary to bridge the gap between low-resolution nighttime light data and high-resolution spatial details. These semantic layers act as mediators that guide the reconstruction process, enabling the model to infer high-resolution spatial patterns from low-resolution input without requiring direct high-resolution nighttime light observations.

The patent creates a reconstructed high-resolution nighttime light image by copying and upsampling the low-resolution input, then refining it through the neural network. The reconstruction model generates a high-resolution version that mimics the appearance and characteristics of true high-resolution nighttime light data, allowing refined applications without actual high-resolution observations.

The reconstruction model creates synthetic high-resolution copies of historical low-resolution nighttime light data. By processing available historical low-resolution data through the trained model, the system generates high-resolution versions that can be used for refined modeling and analysis across the entire historical time series, not just for recent high-resolution observations.

The patent changes the resolution parameter of the nighttime light data through the reconstruction process. The neural network transforms low-resolution input data into high-resolution output by learning the statistical relationships between different resolution levels, effectively changing the spatial scale parameter while preserving the temporal characteristics of the historical data.