Research map / Video / Video segmentation and tracking

Unifying Knowledge from Diverse Datasets to Enhance Spatial-Temporal Modeling: A Granularity-Adaptive Geographical Embedding Approach

ICML 2025 · Video › Video segmentation and tracking · approach: Spatio-temporal predictive modeling

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SQGEF, a Segment Quadtree-based geographical embedding that pretrains on heterogeneous grid and entity CO2 datasets to represent unseen entities at multiple granularities and boosts time-series and spatio-temporal forecasters.

Problem

Spatio-temporal forecasting of geographical entities suffers from limited historical data in scientific datasets, and entities at test time are unseen during training, with heterogeneous granularities and data structures across studies.

Approach

A hierarchical Segment Quadtree over the grid region stores multi-granularity node embeddings, trained via hierarchical grid-based learning (re-divided grids with ancestor-child graphs) and geographical entity-based learning (entities as quadtree node sets), then queried to build interaction graphs for unseen target entities.

Key results

• On China Province, TimesNet w/ Seg reduces MSE from 0.4503 to 0.2743 and MAE from 0.5257 to 0.3900.
• On Europe Country, Informer w/ Seg reduces MSE from 2.0979 to 2.0373 (AGCRN w/ Seg 2.0247 vs 2.2722 baseline).
• Ablation on China Province shows removing Segment Quadtree embedding (w/o SE) or Geographical Entity-based learning (w/o GE) degrades performance for most backbones.

Measured effects reported in the paper

Surprising findings

• On China City, SQGEF slightly degrades performance for two spatio-temporal baselines (GWNet and iTransformer) because under-represented cities are overwhelmed by nearby larger cities.

Limitations

• Noise from under-represented entities surrounded by larger neighbors can hurt performance (China City).
• Experiments limited to carbon-emission datasets with few regions; cross-domain evaluation only briefly reported on air pollution.
• Performance gains are inconsistent across all models (e.g., some spatio-temporal baselines degrade on China City).

Experimental setup

Datasets: ODIAC 1km China ODIAC 1deg China MEIC Model Region Carbonmonitor China ODIAC 1km EU ODIAC 1deg EU Carbonmonitor EU MEIC Model City

Benchmarks: China Province CO2 forecasting Europe Country CO2 forecasting China City CO2 forecasting air pollution datasets (cross-domain)

Metrics: MAE MSE

Baselines: Informer FEDformer Autoformer TimesNet AGCRN MTGNN GWNet TimeXer iTransformer

Builds on: TimesNet MTGNN AGCRN Graph WaveNet

Closest papers by meaning

• Terra: A Multimodal Spatio-Temporal Dataset Spanning the Earth (NeurIPS 2024 · similarity 0.87)
• GeoLLM: Extracting Geospatial Knowledge from Large Language Models (ICLR 2024 · similarity 0.86)
• On the Integration of Spatial-Temporal Knowledge: A Lightweight Approach to Atmospheric Time Series Forecasting (NeurIPS 2025 · similarity 0.84)
• Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.84)
• ST-ReP: Learning Predictive Representations Efficiently for Spatial-Temporal Forecasting (AAAI 2025 · similarity 0.84)
• Poly2Vec: Polymorphic Fourier-Based Encoding of Geospatial Objects for GeoAI Applications (ICML 2025 · similarity 0.84)