High-resolution rural poverty mapping in Pakistan with ensemble deep learning.

Area of Science:

• Environmental science
• Computer science
• Development economics

Background:

• High-resolution poverty mapping is crucial for policy but hindered by data scarcity in many countries.
• Deep learning, particularly Convolutional Neural Networks (CNNs) with satellite data, shows promise for poverty estimation in low- and middle-income countries (LMICs).
• Existing methods often yield coarse spatial resolution, especially in rural areas, limiting their practical application.

Purpose of the Study:

• To develop an accurate and scalable method for high-resolution poverty mapping in rural areas.
• To improve the spatial resolution of poverty estimates beyond current limitations.
• To assess the effectiveness of a transfer learning ensemble approach using CNNs for chronic poverty prediction.

Main Methods:

• A transfer learning approach was used to train three Convolutional Neural Networks (CNNs).
• An ensemble of these CNN models predicted chronic poverty at a 1 km2 resolution in rural Sindh, Pakistan.
• Models were trained using spatially noisy georeferenced household survey data and publicly available satellite imagery (daytime, nighttime) and accessibility data.

Main Results:

• The ensemble model demonstrated superior predictive accuracy in both arid and non-arid regions compared to previous studies.
• Validation exercises, including hold-out, k-fold, and ground-truthing with 7000 households, confirmed the model's reliability.
• The approach achieved higher accuracy metrics than existing methods for poverty mapping.

Conclusions:

• The developed ensemble CNN approach provides an inexpensive and scalable method for high-resolution poverty mapping.
• This technique can significantly enhance poverty targeting and evidence-based policymaking in Pakistan and other LMICs.
• Improved spatial resolution of poverty data is achievable even with limited initial survey data.