Research map / Video / Video segmentation and tracking
Identifying Spatio-Temporal Drivers of Extreme Events
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 an end-to-end model that quantizes multivariate climate variables into binary driver masks (via lookup-free quantization) and predicts extreme agricultural droughts from them, evaluated on three new synthetic benchmarks and two reanalysis datasets.
Problem
Identifying spatio-temporal drivers (anomalies in land-atmosphere and hydrological variables) that precede and are correlated with extreme events such as agricultural droughts, where drivers can be lagged in time and displaced in space, with no driver annotations available.
Approach
Train a Video Swin Transformer feature extractor per variable followed by a lookup-free binary quantization layer so that extremes must be predicted only from binary driver masks; multi-head 3D CNN classifiers and a driver-code loss push drivers to be extreme-correlated.
Key results
- On synthetic CERRA test, the proposed model (Video Swin backbone) reaches 80.44 F1 and 67.28 IoU for driver detection, versus 73.68 F1 for the best baseline ARNet.
- On synthetic NOAA test, the proposed model reaches 82.55 F1 vs 72.02 for DeepMIL.
- Removing the quantization step raises extreme detection F1 only by about 0.96% (synthetic) and 1.93% (EUR-11), suggesting detected drivers carry most predictive information.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed driver-identification method | Synthetic CERRA · Synthetic CERRA test | F1-score: 73.50 -> 80.44 | Helped |
| Use the proposed driver-identification method | Synthetic NOAA · Synthetic NOAA test | F1-score: 72.02 -> 82.55 | Helped |
| Use the proposed driver-identification method | Synthetic artificial data · Synthetic artificial test | F1-score: 61.41 -> 70.33 | Helped |
| Add the quantization loss (ablation) | Proposed method · Synthetic CERRA validation | F1-score (driver): 02.51 -> 82.78 | Helped |
| Add the driver loss (ablation) | Proposed method · Synthetic CERRA validation | F1-score (driver): 29.99 -> 82.78 | Helped |
| Use multi-head extreme prediction instead of one CNN (ablation) | Proposed method · Synthetic CERRA validation | F1-score (driver): 31.23 -> 82.78 | Helped |
| Use independent feature extractors instead of a shared extractor (ablation) | Proposed method · Synthetic CERRA validation | F1-score (driver): 82.39 -> 82.78 | Helped |
| Use temporal attention instead of cross-attention (ablation) | Proposed method · Synthetic CERRA validation | F1-score (driver): 67.18 -> 82.78 | Helped |
| Use LFQ instead of Threshold (Tanh) (ablation) | Proposed method · Synthetic CERRA validation | F1-score (driver): 70.07 -> 82.78 | Helped |
| Use LFQ instead of VQ with Euclidean distance and orthogonality (ablation) | Proposed method · Synthetic CERRA validation | F1-score (driver): 79.26 -> 82.78 | Helped |
| Set temporal resolution T to 6 rather than T=8 (ablation) | Proposed method · Synthetic CERRA validation | F1-score (driver): 77.33 -> 82.78 | Helped |
| Set temporal resolution T to 6 rather than T=1 (ablation) | Proposed method · Synthetic CERRA validation | F1-score (extreme): 69.25 -> 92.45 | Helped |
Surprising findings
- SimpleNet performs strongly on the synthetic CERRA dataset but drops sharply on other synthetic datasets, showing dependence on dataset and backbone.
- Cross-attention between variables degrades performance because anomalies propagate between variables.
Limitations
- Binary driver classification without graded scores; limited temporal window T.
- Does not identify causal relationships; relies on reanalysis data quality.
- Synthetic benchmarks simplify real Earth system dynamics.
Experimental setup
Datasets: ERA5-Land CERRA reanalysis NOAA VHI synthetic CERRA synthetic NOAA synthetic artificial
Benchmarks: synthetic CERRA synthetic NOAA synthetic artificial ERA5-Land EUR-11 drought prediction CERRA real-world drought prediction
Metrics: F1-score IoU overall accuracy
Baselines: Naive Integrated Gradients OCSVM Isolation Forest SimpleNet STEALNet UniAD DeepMIL ARNet RTFM
Builds on: Video Swin Transformer lookup-free quantization vector quantization RTFM integrated gradients
Closest papers by meaning
- Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.85)
- Causal Spatio-Temporal Prediction: An Effective and Efficient Multi-Modal Approach (NeurIPS 2025 · similarity 0.85)
- Terra: A Multimodal Spatio-Temporal Dataset Spanning the Earth (NeurIPS 2024 · similarity 0.84)
- Discovering Latent Causal Graphs from Spatiotemporal Data (ICML 2025 · similarity 0.84)
- DroughtSet: Understanding Drought Through Spatial-Temporal Learning (AAAI 2025 · similarity 0.84)
- How Different from the Past? Spatio-Temporal Time Series Forecasting with Self-Supervised Deviation Learning (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.