Read reference materials

[RemLampa]

https://www.adb.org/sites/default/files/publication/682851/mapping-poverty-satellite-imagery-philippines.pdf

https://www.adb.org/sites/default/files/publication/665961/ewp-629-ai-satellite-imagery-poverty-statistics.pdf

https://github.com/thinkingmachines/ph-poverty-mapping

https://aiforsocialgood.github.io/icml2019/accepted/track1/pdfs/7_aisg_icml2019.pdf

[MoamenAbdelwahed]

@RemLampa What is the definition of done?

[RemLampa]

@MoamenAbdelwahed I believe the intention is for everyone in the task group to be able to read all materials. But let's await the input of the original author.

[josephlaurel]

Based on section 1 of the ADB poverty mapping study:

ADB’s methodology

1. Small area estimation (SAE) of poverty

   • combine survey data with auxiliary information from other data sources, such as census or administrative sources, to produce more granular poverty estimates
   • restrict the explanatory variables in the estimation model to only those that do not vary over time

2. R

3. ML

   • Random forest est. for pop. mapping

4. DL

   • Use of CNN based on Stanford research on combining satellite imagery and machine learning to predict poverty (Jean et al. 2016)
   • data requirements: granular poverty data, daytime satellite images, and nighttime light intensities
   • daytime satellite images as input
   • intensity of night lights used as a proxy for economic activity/development
   • CNN predicts the intensity of night lights
   • transfer learning performed using ImageNet database

5. Ridge regression

   • applied to examine the relationship between the image features as predicting variables and the government-compiled poverty data

6. Inferencing

   • The trained CNN and ridge parameters are then employed to predict poverty using solely a daytime image.