Multi-site deep learning for groundwater level prediction across global datasets: toward scalable applications under data scarcity

ABSTRACT Deep learning (DL), and especially long short-term memory (LSTM) networks, have shown strong potential for groundwater level (GWL) prediction, but remain limited across sites and at scale due to data scarcity and reliance on sparse, inconsistent, or physically ambiguous static site descriptors. This study advances prediction in three key ways. First, we present the first global, large-sample evaluation of multi-site LSTM models trained on over 1,800 wells across nine global regions and evaluate their performance under realistic data constraints. Second, we show that trainable site embeddings enable the model to learn site-specific behavior directly from time series without relying on externally defined site descriptors. Third, we analyze the embedding space, revealing emergent spatial and functional patterns that reflect hydrogeological structure. Embedding-based models achieve strong predictive performance in data-rich regions (median Nash–Sutcliffe efficiency (NSE) > 0.7) and remain robust across a combined global dataset. Our findings further challenge the assumption that larger datasets naturally improve predictions, especially in data-sparse regions. These findings describe both the potential and current limitations of scalable, entity-aware LSTM models for multi-site GWL prediction under data scarcity, paving the way for truly geographically global DL in hydrology.