Earth Observation

Fight Climate Change with AI

Measure

Track greenhouse gas emissions
Estimate carbon footprint
Monitor crops

Mitigate

Target fertilizer and pesticides
Improve energy efficiency
Integrate renewables

Adapt

Detect forest fires
Analyze flood risk
Predict climate conditions

Machine learning projects for earth observation have unique challenges

Diverse Imagery

Remote sensing images can contain a large variety of terrain, seasonal patterns, weather conditions, and lighting changes. Objects of interest may have different characteristics in different parts of the world. This diversity needs to be captured in the training set for an algorithm to properly generalize.

Noisy Images & Labels

Different satellites have different spatial and temporal resolutions. Cloud cover and sensor noise can obscure important parts of the image. Images of the same location from different times may not be perfectly registered. And annotations can change over time. Quality control steps are essential, but imperfect data must also be accommodated.

Sparse Labels

Some applications of computer vision make use of millions of labeled images, whereas smaller data sets are common with more specialized imagery. There may also be a long-tailed distribution with few examples of some classes. Transfer learning and self-supervised methods are often critical to success in this low sample size regime.

Additional Modalities of Data

There are now a number of satellite constellations recording observations not only with visible light but also near infrared, thermal, radar, and more. Combining these modalities with other geographical information obtained on the ground can provide a more powerful means of modeling.

Language Barrier Between Disciplines

Understanding the intricacies of a particular application and possible use cases requires the expertise of domain experts. Project success depends on communicating both ways – about the application and about the machine learning solution.

Build robust and generalizable models

  1. Decipher the distribution shifts in your images
  2. Properly clean your data
  3. Create models that accommodate unique data challenges
  4. Thoroughly validate your models
  5. Iterate and improve

Work With Me