Deep Variational Auto-Encoder for Model-Based Water Quality Patrolling with Intelligent Surface Vehicles

Samuel Yanes Luis, Nicola Basilico, Michele Antonazzi, Daniel Gutiérrez-Reina, Sergio L. Toral Marín

Abstract: This paper addresses persistent monitoring challenges in Lake Ypacaraí, Paraguay, a crucial hydrological resource facing issues of eutrophication and cyanobacteria blooms. Utilizing autonomous surface vehicles equipped with water quality sensors, a model-based approach is proposed for the Non-Homogeneous Informative Patrolling Problem. The UNet based Variational Auto-Encoder architecture is introduced for importance estimation, achieving a 28% and 65% improvement in accuracy for water quality parameters compared to non-parametric approaches such as Gaussian processes and k-Nearest Neighbors, respectively. The proposed model also significantly reduces computational costs, making it suitable for real-time deployment. A greedy patrolling algorithm, exploiting the submodularity of the problem, demonstrates a 41% and 55% performance improvement over algorithms without UNet-VAE. This method enhances monitoring coverage and intensification of high-interest areas, providing a promising approach for hydrological resource surveillance.