Streamflow Synthesis Using an Encoded Textural Pattern Recognition System. II: Model Applications
Shirin Studnicka, Umed Singh Panu · Journal of Hydrologic Engineering · 2025
Pattern recognition-based techniques capture short-term dependencies at the feature extraction stage, whereas long-term dependencies are captured during the formation of feature vectors. The encoded textural feature recognition system developed in Part I of this two-part set of papers introduces a feature extraction approach capable of simultaneously capturing both short-term and long-term dependencies. In this study, the model developed in Part I is applied to synthesize streamflow realizations for three natural watersheds using historical streamflow records. The null hypothesis test conducted on the statistical properties of synthesized realizations and on the historical monthly streamflow of these watersheds confirms that there is no significant difference between the statistical properties of synthesized realizations and the historical monthly streamflow. A comparative analysis between the proposed model and the existing pattern recognition model indicated that the proposed model can preserve the autocorrelation function up to 100 (monthly) lags compared with 24 (monthly) lags in the existing model. Moreover, the Hurst coefficient analysis confirms that the proposed model provides a slightly enhanced representation of statistical characteristics of historical time series, reflecting improved capabilities in modeling long-term dependencies and trends in streamflow data series. Further comparison with artificial neural network (ANN) and autoregressive moving average (ARIMA) models demonstrates that the proposed model effectively captures both the key statistical properties and the seasonal patterns identified by the seasonal index, highlighting its strength in representing temporal structure in hydrological data.