A generalizable and accessible approach to machine learning with global satellite imagery

Area of Science:

• Environmental science
• Socioeconomic studies
• Geospatial analysis

Background:

• Satellite imagery with machine learning (SIML) offers potential for global challenge assessment in data-scarce areas.
• High computational demands of current SIML methods hinder accessibility and widespread application.

Purpose of the Study:

• To develop a computationally efficient SIML method.
• To enable broad accessibility of SIML for researchers globally.
• To demonstrate the generalizability of a single image encoding across diverse tasks.

Main Methods:

• Developed a method using a single, generalized encoding of satellite imagery.
• Achieved competitive accuracy with deep neural networks at significantly lower computational cost.
• Enabled label super-resolution predictions and uncertainty characterization.

Main Results:

• A single satellite image encoding demonstrated generalizability across varied prediction tasks (e.g., forest cover, house prices, road length).
• The method's computational cost was orders of magnitude lower than traditional deep neural networks.
• State-of-the-art SIML performance was achieved by researchers needing only to fit a linear regression.

Conclusions:

• The proposed method democratizes access to advanced SIML by reducing resource requirements.
• Centralized computation and distribution of image encodings facilitate global research collaboration.
• This approach provides a scalable and accessible solution for remote estimation of socioeconomic and environmental conditions.