Pixel-wise Synthetic Hydrological Data for Long-term Modelling: A Novel Approach for Bridging Spatiotemporal Data Gaps

Remote sensing data are crucial for informing earth science models with key hydrological variable, such as evapotranspiration, soil moisture, or terrestrial water storage. However, gaps in historical data, especially pre-2000, hinder long-term hydrological modelling efforts. To address this, we propose a novel method to generate synthetic data, bridging temporal gaps and extending existing datasets for water resource modelling and climate impact studies.

Unlike existing approaches that resample and interpolate historical data as cohesive wholes, the proposed method adopts a pixel-wise perspective. Each pixel’s associated climate time series are analysed independently using a k-Nearest Neighbour (kNN) algorithm paired with a process-specific similarity metric. This allows the identification of pixel-specific analogues based on climate reanalysis data. The selected pixel-wise analogues are then combined to create “compound synthetic images,” preserving spatial and temporal heterogeneity often lost when using a domain-wise approach.

To enhance variability and assess uncertainty, the proposed method integrates stochastic sampling within the analogue selection process. This generates ensembles of synthetic data, enabling quantification of pixel-level uncertainty on any given day.

The proposed approach is tested in the Volta River Basin, a West-African region with strong climate variability and affected by data scarcity. The synthetic data are applied to a spatially distributed hydrological model and evaluated based on their ability to reproduce observed streamflow patterns. Additionally, the model is calibrated separately with real and synthetic data, and the resulting evapotranspiration outputs are compared to assess their closeness.

Preliminary results show that the hydrological model performs equally well in terms of streamflow and evapotranspiration when using either real or synthetic data. This demonstrates the reliability of the synthetic data generation and its suitability for modelling unobserved processes.