Streamflow Synthesis Using an Encoded Textural Pattern Recognition System. I: Model Development
Shirin Studnicka, Umed Singh Panu · Journal of Hydrologic Engineering · 2025
Streamflow synthesis using pattern recognition systems has attracted significant attention in recent years. Feature extraction, a crucial step in this process, enables the identification of various streamflow characteristics. Traditional models rely on immediate past values, thus missing complex relationships across multiple time steps. Given the dynamic and chaotic nature of streamflow, a model capable of capturing both high temporal resolution for immediate past features and low temporal resolution for seasonal patterns is required. This necessitates a paradigm shift of representing streamflow time series in an eight-bit textural image, where two dimensions represent temporal aspects and create a network with intersections forming pixels with gray shade intensity of [0–255] representing the scaled magnitude of streamflow. Texture refers to visual pattern variations in pixel intensities, and thus, features are termed textural features. This study aims to develop a semiautomated model to extract textural features capturing temporal dependencies of each pixel with its previous pixels in horizontal and vertical directions, termed simultaneous autocorrelation. Such two-dimensional correlations are then transformed into the frequency domain using a discrete Fourier transform. A power spectrum of the Fourier coefficients forms transformed textural features, which, in turn, are statistically analyzed to fit a normal distribution. In the synthesis process, transformed textural features are generated within ±1.96 standard deviations of the mean of the fitted normal distribution, and then transformed back to the time domain to synthesize the textural image to decode back into a traditional streamflow time series. The proposed model captures both short- and long-term dependence structures, as demonstrated in Part II of this two-part set of papers.