GPESA: Evolving "Where to Look" in Satellite Imagery for Better Regional Models

TL;DR

Current regional models often rely on "dumb" spatial averaging of satellite data, losing vital information. GPESA (Genetic Programming with Embedded Spatial Aggregation) changes this by treating spatial aggregation as an evolutionary process. It doesn't just learn the weights of a model; it learns the location and size of the geographical regions that matter most. In predicting snow properties in High-Mountain Asia, this method outperformed state-of-the-art Lasso regression by a significant margin.

Problem & Motivation: The Up-sampling Trap

In remote sensing, we are blessed with high-resolution grids (pixels), but the goals are often regional (e.g., "how much total water is in this mountain range?"). Researchers usually solve this in two flawed ways:

1. Uniform Up-sampling: Averaging everything. This is too simplistic; it treats a shadow-filled valley the same as a snowy peak.
2. Pixel-as-Feature: Treating every pixel as a variable. This leads to the "curse of dimensionality," where you have thousands of features but very few historical data points, causing models to overfit and fail on new data.

The authors hypothesized that there are specific, non-uniform spatial subsets that hold the most predictive power. The challenge was finding them without manual "guesswork."

Methodology: Evolution as a Spatial Architect

The core innovation is GPESA. Instead of using fixed inputs, the model uses "Embedded Spatial Aggregations."

1. The Dynamic Terminal

In a standard Genetic Programming (GP) tree, a terminal might be a variable like "Mean Snow Cover." In GPESA, a terminal is a function: Average(Variable, X, Y, Radius).

• X, Y: The center of a circular region.
• Radius: The extent of the aggregation.

2. Simultaneous Evolution

As the GP evolves the mathematical formula (e.g., $Aimes\sin (B)$), it also "mutates" the coordinates and radii of these circles. This means the model finds the optimal geometry of the features at the same time it builds the regression equation.

Fig 1: In GPESA, terminals are parameterized circles ($x,y,r$) that evolve alongside the tree structure.

3. Optimization Objectives

To prevent the models from becoming overly complex (bloat), the authors used Age-Fitness Pareto Optimization (AFPO). They optimized for three things:

• Error: How accurate is the prediction?
• Size: How small is the formula? (Parsimony)
• Age: How long has this lineage existed? (To preserve diversity)

Experiments & Results: Crushing the Baselines

The authors tested their method on a 1935-day dataset of Snow Water Equivalent (SWE) in High-Mountain Asia. They compared GPESA against:

• Standard Lasso (SL): The industry standard for sparse linear modeling.
• Ridge Regression: A standard linear baseline.
• Filter/Wrapper Methods: Semi-automated ways of selecting regions.

GPESA not only had the lowest mean error but was statistically superior to Lasso in a majority of the test years.

Fig 2: Heatmaps showing which spatial units the different models deemed "important." GPESA (f) shows a more nuanced, evolved understanding of relevant geographical zones compared to the rigid grids of filter methods.

Critical Analysis & Conclusion

The "White Box" Advantage

Unlike Deep Learning, where the decision-making process is hidden, GPESA produces human-readable formulas and clear geographical maps of what it "thinks" is important. By looking at the evolved circles, researchers can identify specific mountain corridors that act as bellwethers for regional snow levels.

Limitations

The main drawback is computational cost. The "on-the-fly" aggregation requires calculating means within varying circles thousands of times per generation. The authors noted a 500% increase in training time compared to standard GP.

Looking Forward

GPESA represents a shift toward physically-grounded AI. Instead of forcing data into a fixed grid, we allow the AI to define its own spatial context. Future iterations could involve evolving complex shapes (not just circles) or integrating temporal windows into the evolution, allowing the model to decide both when and where to look.

Takeaway: If your spatial data is too high-dimensional for standard regression but you need more interpretability than a Black-Box CNN, evolving your features via GPESA is a powerful middle ground.