Textural Image-Based Feature Prediction Model for Stochastic Streamflow Synthesis

Abstract

Abstract To address the challenge of obtaining reliable streamflow data for water resource management, this paper develops an encoding scheme to transform a streamflow time series into an 8-bit grayscale image where it is feasible to develop a correlation between neighbouring pixels to reveal patterns that are not easily discernable in traditional time series analysis. To extract relevant information from an encoded streamflow image, a novel textural feature extraction approach has been developed for synthesizing streamflow data. The developed textural feature extraction model can capture the simultaneous correlation in two dimensions of an encoded streamflow image, which is then transformed into the frequency domain using a discrete Fourier transform. The use of the power spectrum of the Fourier coefficients facilitates the synthesis of encoded streamflow images. The effectiveness of the model is evaluated using three case studies across Canada by comparing the properties of synthesized streamflow with the historical streamflow using the structural similarity (SSIM) index. Results show that the proposed model effectively synthesizes encoded streamflow with high SSIM values for the Fraser, Black, and South Saskatchewan Rivers. The results also affirm that the model reproduces the temporal dependence and correlation structure of the historical streamflow and the average of 30 synthesized realizations up to more than 100 lags. The null hypothesis tests support the conclusion that there is statistically no significant difference between the synthesized monthly streamflow time series and the historical time series. In addition, the entropy-based test also emphasizes that synthesized and historical streamflow are indistinguishable.