Using Artificial Intelligence and Satellite Imagery to Zero In on the Philippines’ Most Vulnerable Communities
summary
Claims that machine learning on daytime satellite tiles plus nighttime lights proxy can reconstruct poverty at finer scales, achieving cross‑validated R² ≈ 0.63 against DHS cluster wealth indices. Pipeline uses CNN feature embeddings trained on Google static maps to predict light intensity, averaged per cluster and fed to ridge regression; results identify unexpected poor pockets (e.g., a barangay at 76% poverty) and map conflict‑area poverty. Geospatial work moves from QGIS to GeoPandas + Python for reproducibility and production.
AI-written summary · about
chapters (10)
- 0:00Start
- 8:06Predict the Corresponding Nighttime Light Intensity
- 8:30Convolutional Neural Networks
- 11:11Results
- 14:10High-Resolution Mapping for Conflict Areas
- 15:23Limitations
- 22:55Final Thoughts
- 23:42Use Case from Poverty Mapping
- 34:15Most of the Cost Comes from Downloading Satellite Images
- 37:52Cross Validation
related talks
37mSpatial Data Analysis With Python
07mMapping Urban Deprivation
22mIdentifying Urban Makeshift Communities using satellite imagery and geo-coded data
58mImproving disaster response with machine learning
36mGeoMonitoring with Python: Innovative Solutions for Latin America’s Challenges
30mUsing AI to Spot Deforestation-related Cows on Satellite Images
21mNeedle in the Haystack: Applied ML for Inspections of Power Infrastructure
27mThe View from Above Getting Started with Satellite Images