SIGSPATIAL 2026: satellite VLMs recover 98.3% of travel-census accuracy; AlphaEarth wins cross-continental transfer by 33%
flosalim · x · 2026-10-20
The author's lab landed 9 papers at SIGSPATIAL 2026; two tackle urban mobility from orbit.
- Paper 1 (Do Satellites See Commuters?) benchmarks vision encoders (RemoteCLIP, DINOv3, SatCLIP, AlphaEarth) via a WeDAN graph diffusion pipeline across 1,925 US counties, 325 UK districts, and 14 cities, asking whether satellite foundation models can replace sparse, costly travel censuses.
- Finding 1: language-supervised models (RemoteCLIP) dominate locally, recovering 98.3% of census performance; geographically grounded models (AlphaEarth) win cross-continental transfer by 33%.
- Finding 2: scale isn't everything. The attached figure also outlines the open-source Massive-STEPS dataset and benchmark — supervised/zero-shot next-POI recommendation and spatiotemporal reasoning tasks, evaluated from classical sequential models to LLM-based methods (LLM-Mob, LLM-ZS, LLM-Move), with longitudinal and temporal-shift sensitivity analyses.
Related event: Satellite Imagery with VLMs Can Recover 98.3% of Commute Survey Data(2 posts)→
More from Research
- Walden Robotics lets robots 'phone a friend' mid-task, turning human help into training data — siddkaramcheti · 2026-10-20
- LLM Agents Shred Security Through Obscurity: Copier Currency Detection Reversed in Days — moyix · 2026-10-20
- Sakana AI's QD-Rec Applies Quality-Diversity Evolutionary Search to Open-Ended Recommendation — SakanaAILabs · 2026-10-20
- POPE benchmark audit: yes/no readout choice alone shifts VLM F1 by 6.13 points — Kesav Kumar Jayakumar · 2026-10-20
- Meta's synthetic data pipeline puts AI agents in simulated 'digital lives' that mimic real apps — Puzzleheaded-King584 · 2026-10-20
- typed-decisions is now an official HF benchmark ranking 2,274 open-source decision models — LysandreJik · 2026-10-20