Streambed water flux characterization through a Deep-Learning-Based approach considering data worth analysis: Numerical modeling and sandbox experiments

Abstract

• An approach named CNN-DWA is proposed for estimating streambed water fluxes. • High-utility data are chosen through data worth analysis for CNN model training. • A few chosen data points provide similar information content to entire dataset. • CNN-DWA achieves more accurate flux estimations than the previously used IES. Quantifying streambed water fluxes (SWFs) is a prerequisite for studying the transport and fate of contaminants in the hyporheic zone. One approach, the heat tracing technique combined with analytical or numerical heat transport models has been widely used to quantify SWFs. However, both analytical and numerical methods are limited. For example, analytical methods are built upon rarely met assumptions, while the inversion performance of numerical methods can be impeded by uncertainties from different sources. To overcome these limitations, we introduce a new approach that combines convolutional neural networks (CNNs) with the Bayesian data worth analysis (DWA) framework for estimating SWFs. In this approach, a synthetic dataset is first generated through numerical simulations of groundwater flow and heat transport in streambeds. Then high-utility temperatures are chosen from the synthetic dataset through Bayesian DWA and further used to train the CNN model. We assess the capability of Bayesian DWA in improving the SWF estimation accuracy of CNN models through both numerical and sandbox experimental cases. Results show that the CNN model trained on temperatures with higher utility generally gives more accurate SWF estimations; use of a few high-utility temperature measurements can provide near similar constraint on the CNN model training and SWF estimation as compared to the full time series. Furthermore, a comparison to the iterative ensemble smoother algorithm shows the proposed inversion approach can estimate the flux field in the sandbox more accurately. The proposed CNN-based inversion approach considering DWA (CNN-DWA) would be beneficial more broadly for other hydrological problems.