Integrating remote sensing and unsupervised machine learning techniques for runoff prediction in ungauged catchments through spatiotemporal catchment similarity analysis

Abstract

Predicting streamflow in ungauged basins is a significant challenge in hydrology. This study presents a framework that integrates remote sensing and machine learning to predict runoff for ungauged catchments by analyzing catchment similarity and evaluating whether training on hydrologically similar catchments outperforms training on all available streamflow and precipitation data across catchments. The study analyzed 30 catchments across Sri Lanka's wet, intermediate, and dry zones using two decades of streamflow and precipitation data. An initial suite of 14 hydrological signatures derived from daily river discharge and rainfall time series was refined via correlation analysis (|r| ≥ 0.7) into six non-redundant signatures (Runoff Ratio, Baseflow Index, Flow Duration Curve Slope, Recession Constant, Precipitation Elasticity, Rising Limb Density) that together represent the complete unit hydrograph. For classification, multiple algorithms (K- Means, K-Medoids, Hierarchical) and distance metrics (Euclidean, Manhattan, Cosine) were evaluated. Hierarchical clustering with cosine distance proved optimal, forming seven distinct clusters with a high Silhouette Score of 0.525. Additional validation using satellite-derived data confirmed strong hydrological coherence (Hydrological Coherence Index = 1.48). For ungauged applications, signatures were predicted from satellite data, achieving a pairwise accuracy of 87.4% in replicating cluster relationships, although exact membership accuracy was lower (60%) due to borderline catchments. The prediction compared three approaches for pseudo-ungauged basins: Monaragala (dry zone) and Ellagawa (wet zone). In both the Dry and Wet zones, the traditional HEC-Hydrologic Modeling System parameter transfer failed to deliver sufficient predictions (Monaragala: Nash-Sutcliffe Efficiency (NSE) = -0.407; Ellagawa: NSE = 0.425). Long Short- Term Memory (LSTM) models were universally superior, but the optimal strategy was regime dependent. In the Wet Zone, the Regionalized LSTM (trained on similar donors) performed best (Ellagawa: NSE = 0.709, Root Mean Square Error (RMSE) = 4.20 mm/day). Conversely, in the Dry Zone, the Global LSTM (trained on all 30 catchments) was more effective (Monaragala: NSE = 0.535, RMSE = 2.88 mm/day), suggesting that data volume aids learning of complex arid hydrology, though low-flow prediction remained a challenge for all models. The findings mandate a regime-specific prediction strategy: hydrological similarity guides optimal donor selection in wet zones, while a global data approach is more effective in dry zones.