maryadesoba01-wq/Deep-Learning-Based-Poverty-Estimation-in-Nigeria-Using-Satellite-Imagery-

# Deep Learning-Based Poverty Estimation in Nigeria **Deep Learning-Based Poverty Estimation in Nigeria Using Sentinel-2 Satellite Imagery and DHS Socioeconomic Data** This project implements a modular, end-to-end deep learning framework to estimate household wealth indices across Nigeria. By integrating household-level socioeconomic survey records from the **Demographic and Health Survey (DHS) 2018** with high-resolution earth observation data (Sentinel-2 multi-spectral bands), it estimates local wealth based on satellite features. The pipeline implements regression models (Baseline CNN and ResNet50 Transfer Learning) to predict numerical wealth targets and features Grad-CAM-based Explainable AI (XAI) to ensure model interpretability. ## Key Features - **Data Integration Pipeline:** Merges DHS survey socio-economic target data (`hv271`) with GPS cluster centroids. - **Satellite Imagery Extraction:** Capable of downloading Sentinel-2 ($224 \times 224$ patches) from Google Earth Engine or processing large local GeoTIFF files. - **Deep Learning Architectures:** Supports training a multi-layer Convolutional Neural Network (CNN) and fine-tuning a pre-trained ResNet50. - **Explainable AI (XAI):** Generates Grad-CAM heatmaps overlayed on RGB satellite inputs to illustrate the geographical features influencing poverty estimates. - **Automated Validation & Evaluation:** Splits data, assesses models using RMSE, MAE, and $R^2$ metrics, and outputs comprehensive scatter plots and logs. ## Project Structure ```text mary phd/ ├── data/ # Generated datasets (e.g., extracted satellite patches, processed CSVs) ├── figures/ # Generated visualizations (training curves, Grad-CAM overlays, plots) ├── models/ # Stored model weights and architecture bins (`.keras`) ├── outputs/ # Output predictions and CSV splits ├── results/ # Metrics, evaluations, and summary reports ├── src/ …