A deep surrogate model with spatio-temporal awareness for water quality sensor measurement
Yifan Zhang, Peter J. Thorburn
Abstract
Measuring some key variables on a large scale is neither practical nor financially viable in agricultural and industrial systems. It is possible to estimate the key variable based on other simultaneously measured variables. However, when applying the surrogate model to a new location, the surrogate performance could decrease significantly. We propose a deep surrogate model (DSM) with spatio-temporal awareness for estimating water quality variables. The DSM uses a stacked denoising autoencoder to extract the features of raw sensor data and encodes the temporal and auxiliary information to improve the generalization of the DSM. The domain adaptation layer is designed to learn the spatial differences between monitoring stations in disparate locations. The experimental results indicate that the DSM outperforms five alternative methods in generating estimated nitrate concentration. Accordingly, the DSM is an encouraging approach for estimating water quality constituents in large-scale water quality monitoring networks.