wenxinw81/AfriNTL-code

A 500 m annual nighttime light dataset optimized for Africa from 2000 to 2025 # AfriNTL Code AfriNTL is an Africa-optimized two-stage nighttime light reconstruction framework for producing annual 500 m nighttime light data. The implementation follows the manuscript methods: Stage I calibrates and desaturates DMSP-OLS time series with a Dynamic Spatio-Temporal Feature Extraction Network (DSTFE), and Stage II reconstructs high-resolution 500 m nighttime lights with population-adaptive sparse convolution and an enhancement-denoising dual-branch network. ## Method Overview Stage I uses a seven-year temporal window from `Y-3` to `Y+3`, year-difference vectors, saturation-mask vectors, a UNet encoder-decoder, and multi-head temporal attention to generate temporally consistent 1 km calibrated nighttime light fields. Stage II integrates Stage I output, LandScan population density, DEM-derived terrain relief, NDVI, and MODIS/SDC500 background features. It adapts the receptive field between sparse rural settlements and dense urban agglomerations, enhances weak valid light signals, suppresses geographically induced noise, and outputs 500 m reconstructed radiance. ## Repository Layout ```text AfriNTL/ models/ PyTorch Stage I and Stage II model definitions data/ GeoTIFF IO, temporal windows, and dataset classes training/ L1-loss training and evaluation loops inference/ Annual raster prediction helpers metrics/ Radiometric, temporal, super-resolution, and urban metrics configs/ Default YAML configuration scripts/ Training, prediction, and validation entry points examples/ Synthetic forward-pass example tests/ Synthetic unit tests ``` ## Installation ```bash python -m venv .venv source .venv/bin/activate pip install -e ".[dev]" ``` ## Configuration Edit `configs/AfriNTL_default.yaml` to match the local project layout. The default configuration follows the manuscript data groups: - harmonized DMSP-OLS annual GeoTIFFs - RNTL calibration reference GeoTIFFs - VIIRS VNL V2 annual GeoTIFFs - LandSc …