Research map / Time series / Causal and generative time series
DroughtSet: Understanding Drought Through Spatial-Temporal Learning
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.
Introduces the DroughtSet CONUS drought benchmark (2003-2013 weekly remote-sensing and reanalysis data) and SPDrought, a spatial-temporal Transformer that jointly forecasts three drought indices 26 weeks ahead and interprets them with Integrated Gradients.
Problem
Subseasonal-to-seasonal (S2S) drought prediction is hard because physical and biological drivers interact, and existing deep learning work covers few drought types and ignores spatial neighbors and joint prediction.
Approach
A spatial-temporal model that uses static-feature attention over a 5x5 neighborhood to aggregate neighbors' dynamic features, encodes static and dynamic features with separate networks feeding a Transformer encoder-decoder, and uses shared multi-task regressors for three drought indices, plus Integrated Gradients for attribution.
Key results
- SPDrought achieves total MAE 38.01 (x10^-3) over soil moisture, ESI, and SIF, versus 45.18 for the best baseline (TimesNet) and 75.41 for DLinear.
- On soil moisture drought classification, SPDrought reaches 86.26% accuracy and 76.80% precision, versus 81.54% accuracy for TimesNet.
- Multi-task SPDrought roughly matches single-task variants (total MAE 38.01 vs 33.24) while cutting training time to about one-third.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPDrought instead of the strongest reported baseline, TimesNet | DroughtSet drought-index forecasting · Soil Moisture | MAE (× 10⁻³): 25.96 -> 21.39 | Helped |
| Use SPDrought instead of the strongest reported baseline, TimesNet | DroughtSet drought-index forecasting · ESI | MAE (× 10⁻³): 5.11 -> 4.40 | Helped |
| Use SPDrought instead of the strongest reported baseline, TimesNet | DroughtSet drought-index forecasting · SIF | MAE (× 10⁻³): 14.11 -> 12.21 | Helped |
| Use SPDrought instead of the strongest reported baseline, TimesNet | DroughtSet drought-index forecasting · Total | MAE (× 10⁻³): 45.18 -> 38.01 | Helped |
| Use SPDrought instead of the strongest reported baseline, TimesNet | soil-moisture percentile drought prediction · Soil moisture drought — test set | Accuracy: 81.54 -> 86.26 | Helped |
| Use SPDrought instead of the strongest reported baseline, TimesNet | soil-moisture percentile drought prediction · Soil moisture drought — test set | Precision: 68.98 -> 76.80 | Helped |
| Add multi-task joint training instead of training separate single-task models (ablation) | SPDrought · DroughtSet — Soil Moisture | MAE (× 10⁻³): 19.39 -> 21.39 | Hurt |
| Add multi-task joint training instead of training separate single-task models (ablation) | SPDrought · DroughtSet — ESI | MAE (× 10⁻³): 3.35 -> 4.40 | Hurt |
| Add multi-task joint training instead of training separate single-task models (ablation) | SPDrought · DroughtSet — SIF | MAE (× 10⁻³): 10.50 -> 12.21 | Hurt |
| Add Transformer encoder (ablation) | SPDrought · DroughtSet — Total | MAE (× 10⁻³): 50.95 -> 38.01 | Helped |
| Add Transformer decoder (ablation) | SPDrought · DroughtSet — Total | MAE (× 10⁻³): 41.01 -> 38.01 | Helped |
| Use the full training window rather than reducing it to 50 weeks (ablation) | SPDrought · DroughtSet — Total | MAE (× 10⁻³): 39.27 -> 38.01 | Helped |
Surprising findings
- Multi-task training is slightly worse in MAE than single-task training (38.01 vs 33.24 total), though the authors frame it as comparable and faster.
- DLinear, which is linear, performs worst on DroughtSet (total MAE 75.41), contrary to its reported strength on other forecasting benchmarks.
Limitations
- Study limited to CONUS, so the trained model is not directly deployable to other regions.
- Temporal interplay among dynamic predictors is not analyzed; only spatial influence is examined.
- Dataset release is stated only as occurring upon acceptance, and the paper provides a GitHub link for resources.
Experimental setup
Datasets: NLDAS ALEXI GridMET CSIF ERA5 SRTM VODCA MODIS LAI GLAD NLCD
Benchmarks: DroughtSet
Metrics: MAE accuracy precision p-value (paired t-test)
Baselines: Transformer Informer PatchTST DLinear iTransformer TimesNet LSTM
Builds on: Transformer Integrated Gradients DLinear PatchTST
Closest papers by meaning
- SIDE: Socially Informed Drought Estimation Toward Understanding Societal Impact Dynamics of Environmental Crisis (AAAI 2025 · similarity 0.87)
- Terra: A Multimodal Spatio-Temporal Dataset Spanning the Earth (NeurIPS 2024 · similarity 0.84)
- Identifying Spatio-Temporal Drivers of Extreme Events (NeurIPS 2024 · similarity 0.84)
- Multi-Scale Graph Learning for Anti-Sparse Downscaling (AAAI 2025 · similarity 0.83)
- Quantile-Regression-Ensemble: A Deep Learning Algorithm for Downscaling Extreme Precipitation (AAAI 2024 · similarity 0.83)
- Multi-modal Learning for Geospatial Vegetation Forecasting (CVPR 2024 · similarity 0.83)
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.