Small Enough to Matter: Distilling Geospatial Foundation Models into Climate-Ready Digital Twins

Introduction

Earth observation AI has grown enormous, and that growth quietly excludes the places most exposed to climate stress. Communities tracking drought, land degradation, or seasonal flooding rarely have a GPU cluster on hand. GeoFM-DistillTwin asks a plainer question: can a satellite-trained model shrink to something a modest machine can run, without losing the reasoning that makes it useful? The answer sits close to three Sustainable Development Goals: Climate Action, Industry Innovation and Infrastructure, and Life on Land.

Methods

We start from a large pretrained geospatial foundation model as teacher and train a far smaller student to match its spectral and temporal reasoning, not merely its output labels, through a spectrotemporal alignment loss. Because satellite revisits arrive unevenly, a gap-aware, time-conditioned preprocessing step teaches the student to reason across irregular intervals rather than a fixed weekly rhythm. A lightweight synchronization rule then updates the student model of the landscape only where meaningful change is detected, instead of recomputing everything each pass.

Results

The distilled student needed twenty-two times fewer parameters than its teacher and ran about ten times faster on ordinary CPU hardware, while its segmentation accuracy, an IoU of 0.647, exceeded the teacher IoU of 0.561. Error against a naive baseline fell by a quarter, and selective updating cut redundant recomputation by nearly ninety-five percent. A small simulation shows the same pattern: a synthetic student compressed twenty-two fold recovers 0.305 IoU through distillation alone, moving from 0.456 to 0.761 agreement with its teacher, with no extra data or hardware.

Conclusions

A model that fits on modest hardware can still see the land clearly enough to matter. That is the premise behind GeoFM-DistillTwin: keep the reasoning that makes geospatial AI useful, shed the weight that only makes it expensive, and let climate monitoring stop depending on who can afford a data centre.