Subsurface Insight: Spatio-Temporal Embeddings for Hydrogeological Time Series Analysis

Quach Van Dang, Lynh HOANG-Vy-Thuy, Tuan-Anh Tran

Keywords: Time Series Analysis, Groundwater Level, Spatio-Temporal Embeddings, Hydrogeological Time Serie, Unsupervised Clustering

TL;DR: Our spatio-temporal model uses contrastive learning on groundwater data to generate deep embeddings that strongly align with expert hydrological classifications (NMI > 0.55), proving the capture of geology meaningful subsurface characteristics.

Abstract: The generation of terabytes of hydrogeological sensor data each year has driven a growing demand for hydrological data analysis to aid in water resource prediction and management. However, these data present complex spatio-temporal dependencies, especially for large-scale data. Traditional statistical methods like PCA often fail to capture non-linear temporal patterns, and existing deep learning approaches do not effectively focus on integrating spatial context. This paper introduces a deep spatio-temporal autoencoder architecture to learn embeddings from decades of French groundwater level data. By using contrastive learning, we combine time series and their geographical coordinates to generate low-dimensional embeddings. We demonstrate that these embedding vectors are highly effective for downstream tasks, unsupervised clustering, compared with traditional methods. Crucially, our approach achieves a Normalized Mutual Information (NMI) score exceeding 0.55 against an expert-labeled ground truth, confirming that the learned representations capture physically meaningful subsurface characteristics.

Serve As Reviewer: ~Quach_Van_Dang1

Submission Number: 29