Predicting poverty

Approach 1 - Nightitme Lights

Research by J Neal et al.

  • Use Nighttime light values (NLV) as a proxy for poverty
  • Dataset provided by Earth Observation Group and Images from Google Static Map at 2.5 Meters of resolution
  • Using nightlights as a proxy for development. Nightlights data cannot be directly used because there is less difference in luminosity between rich and poor regions in Africa
  • Use Day images which capture more information and use night images as label data.
  • Use transfer learning to learn features from day satellite images and NLV labels
  • Train the model and learn important features. These features are calculated for a new image.
  • Use these learned features along with survey data at cluster level to train the model.
  • As the cluster level data is very less, use simple models
  • Use the trained model for classifying new clusters or areas

Using CNN on Nightlight Images to learn features

Approach 2 - DHS Data & Phone CDR

Research by Prof.Joshua Blumenstock et al.

  • Publicly available satellite-based estimates of poverty are available
  • The estimation methods use deep learning models trained on Demographic and Health Surveys (DHS) data from neighbouring countries to estimate the average relative wealth of each 2.4km tile in Togo
  • This is used to do cluster level predictions
  • Identify the clusters
  • Use phone CDR as independent variables and survey data as dependent variable to build models for each household

Using Phone CDR and Survey data for prediction

Approach 3 - Land Use

Research by Ayush et al 2020

  • Land use and the manufactured objects observed in a satellite image emphasize the wealthiness of an area
  • CNN was trained on a land use detection and classification task
  • They used xView data consisting of very high resolution images annotated with bounding boxes defined over 10 main classes (building, fixed-wing aircraft, passenger vehicle, truck, railway vehicle, maritime vessel, engineering vehicle, helipad, vehicle lot, construction site) and 60 sub-classes.
  • Yolo V3 was used for object detection

Approach 4 - Tile2Vec

Research by Jean et al 2019

  • Based on Unsupervised learning
  • This method emphasizes the difference between two satellite images
  • Cluster homogeneous-looking areas and assume that some clusters will be specific to poor areas

Fields are often surronded by other fields

Contrastive learning between Anchor Neighbor and Distant tiles

Comparison of Different Approaches

comparison of different approaches