Satellite imagery of a densely built settlement

IEEE Research Paper · Co-Author

Harnessing Satellite Imagery with Convolutional Neural Networks for Poverty Prediction

2024 International Conference on Artificial Intelligence and Quantum Computation-Based Sensor Application (ICAIQSA) · December 2024

PythonPyTorchCNNComputer Vision

📍 Presented at the 2024 International Conference on Artificial Intelligence and Quantum Computation-Based Sensor Application (ICAIQSA), December 2024.

Ground-truth poverty data is expensive and slow to collect — it usually means door-to-door household surveys across regions that are often hard to reach. This paper asks a different question: can a computer look at a satellite photograph and estimate how poor an area is, without anyone ever setting foot there?

The approach uses a convolutional neural network (CNN) — the same class of model that powers modern image recognition — trained on satellite imagery of Africa. Rather than being told explicitly what to look for, the network learns on its own which visual patterns correlate with poverty: road density, the layout and materials of buildings, vegetation cover, and general land use.

Satellite imagery of settlement adjacent to green space — Dense settlement bordering open land — the kind of texture and layout the model learns from.

Satellite imagery of settlement with unpaved roads — Roof material, road paving, and building density all carry signal about an area's economic conditions.

To make sure the model was actually learning something generalizable, the dataset was split 70% for training and 30% for testing. The model never saw the test images during training, so its performance on that held-out set is a fair measure of how well it would work on a brand-new region.

Methodology flowchart: image dataset to data visualization, preprocessing, conversion, splitting, training, and result — Fig. 1 — the full pipeline: raw image dataset → preprocessing → CNN training → evaluation on held-out test data.

The final model reached 90% accuracy on the test set — a strong result for a problem this indirect, where the model has to infer an economic condition purely from visual texture and geometry.

Satellite imagery of a rural village settlement — A more rural settlement pattern from the same imagery set — sparser buildings, different road structure.

The paper was co-authored and published to IEEE Xplore.

View on IEEE Xplore

📖 Cited By

Hen Hen Lukmana, Muhammad Al-Husaini, Iseu Siti Aisyah, Lilik Hidayanti, Luh Desi Puspareni, Vito Hafizh Cahaya Putra, Muhammad Sidiq Asyaky, “Edge-Aware AI-Model: A Lightweight Deep-Learning for Mapping of Child Growth,” 2025 Tenth International Conference on Informatics and Computing (ICIC), pp. 1–5, 2025.

View Citing Paper on IEEE Xplore