RoSriv30/settlements-without-electricity

Using neural networks (ResNet, UNet, UNet++) to detect settlements that lack electricity in Sub-Saharan Africa # Identifying Settlements Without Electricity ## Overview ### Challenge - The need for this research stems from a critical development challenge: over 600 million people in Africa lack access to electricity, hindering economic development and contributing to persistent poverty and inequality. Traditional methods of identifying unelectrified regions are often resource-intensive, requiring extensive ground surveys and infrastructure assessments. Satellite imagery, with its ability to cover vast areas efficiently, presents an innovative solution to this problem. However, the complexity of satellite data, including variations in resolution, spectral bands, and temporal coverage across different satellite systems, poses significant analytical challenges. There is a pressing need for robust methodologies that can accurately process and interpret this data to identify non-electrified settlements. By addressing this need, this project aims to support electrification planning and implementation, contributing to the broader goal of universal access to electricity and the United Nations Sustainable Development Goals. - The source of the original satellite dataset used in this project is the 2021 IEEE GRSS Data Fusion Contest: Track DSE. The dataset comprises 60 folders, each representing a distinct 64 km^2 geographical region. Within each folder, there exists a collection of 98 images, corresponding to various bands captured by four different satellites: Sentinel-1, Sentinel-2, Landsat 8, and the Suomi VIIRS night-time dataset. These images are standardized to a resolution of 800×800 pixels. Each band captures different aspects of the Earth's surface, ranging from intensity values for VV and VH polarization (Sentinel-1) to reflectance data across various spectral ranges (Sentinel-2, Landsat 8). Link To the Dataset. ### Preprocessing - Functions were written to extract channels from images, including Sentinel-1, Sentinel-2, Landsat 8, and the Suomi VIIRS. Metadata …