Quality over Quantity: Number of Samples Required to Train Suspended Sediment Concentration Soft Sensor Models is Reduced Using Gaussian Variance Reduction via Orthogonal Matching Pursuit

Abstract

The deployment of soft sensors, which estimate difficult-to-measure parameters from surrogate data, is often hindered in environmental monitoring by the high cost of field samples and site-specific nature of model training. Even when trained at one location, scaling soft sensor models to new sites requires extensive, additional retraining efforts. This paper addresses this challenge by applying a methodology to transfer knowledge from a transfer knowledge from a well-characterized soft sensor model at an existing location to guide efficient data acquisition for suspended sediment soft sensor development at new locations. Our approach leverages Gaussian process (GP) regression, using the information encoded in a site’s GP hyperparameters to inform sample selection. Specifically, we employ orthogonal matching pursuit as an algorithmic mechanism for Gaussian variance reduction, selecting a minimal yet highly informative set of training samples to maximize the new model’s performance. Evaluating this methodology on acoustic and suspended sediment data at ten river locations using one of two sensor types, we demonstrate its potential to reduce the number of samples required while achieving comparable model accuracy to regression models trained on larger data sets. We present a comparison of active learning vs static transfer of knowledge between sites and a practical deployment implementation of this strategy.