Spectrum Transformation Enhanced Spatiotemporal Learning for Decadal Hydrological Forecasts

Decadal hydrological prediction underpins strategic decisions in water security, infrastructure investment, and disaster risk reduction by extending actionable guidance beyond seasonal horizons. Large-scale climate indices often exhibit longer predictability than regional precipitation itself and leveraging them into multiyear hydrological prediction can be more effective than relying directly on raw decadal forecasts. Still, decadal predictability is constrained by two barriers: (1) scarce training samples in decadal climate prediction project (DCPP) and (2) spectral mismatch between climate predictors and hydrological responses. Although machine learning (ML) models have high capacity, independent grid-lead training routinely overfits DCPP forecasts and therefore fails to outperform regression model baselines. We overcome these limits by coupling spectral alignment using wavelet prediction system (WASP) with a spatiotemporal merging architecture. WASP decomposes relevant climate predictors into multi-scale components and learns frequency-targeted weights to align predictor spectra with local hydrological responses. Spatiotemporal merging then pools information across space and leads, expanding the effective sample size, stabilizing complex learners, and promoting spatiotemporally coherent outlooks.

Applied to Australian drought forecasting, the framework systematically shows an increasing prediction skill in 87% of grids with a mean gain of 0.16 in correlation relative to the regression model. Event-based diagnostics show more faithful results of extreme events, including the 2002 Millennium Drought and wet spells around 2001. This method also skilfully forecasts the prolonged 2018–2020 Australian drought.

Our results elucidate the critical dependency of decadal drought prediction skill on the interplay between model complexity and predictor quality: Spatiotemporal pooling stabilizes training in complex models and improves generalization instead of overfitting when trained independently. Crucially, we identify a predictability horizon beyond 36 months where skill peaks and the advantage of WASP over raw predictors vanishes, indicating that decadal forecast quality is limited by the performance of the underlying dynamical climate models rather than by post-processing techniques. These advances provide practical value for agencies such as WaterNSW in Australia, offering scientific guidance for reservoir operation, integrated water resources planning and climate resilient adaptation strategies for national communities and ecosystems.