Environmental Monitoring in New Zealand: Exploring Spatiotemporal Relationships

Marcos Vinícius dos Santos Ferreira, Luiz Cláudio Dantas Cavalcanti, Tatiane Nogueira Rios, Guilherme Weigert Cassales, Nick Jin Sean Lim, Albert Bifet, RICARDO RIOS

Keywords: Environmental Monitoring, Spatiotemporal Modeling, Foundation Models, Graph Neural Networks

TL;DR: This study compares statistical, recurrent, foundation, and graph-based AI models for temperature forecasting across New Zealand.

Abstract: The accelerating pace of global warming, confirmed by the Intergovernmental Panel on Climate Change (IPCC), underscores the urgent need for continuous environmental monitoring and adaptive modeling frameworks. Therefore, this study explores AI-driven temperature forecasting across New Zealand using real observations from the national MetService network within the TAIAO data science programme. We evaluate a diverse set of models, from statistical baselines to deep neural networks, foundation models, and graph-based architectures, to assess their capacity for adaptive, spatially aware prediction. Our results show that graph-based representations substantially improve the modeling of spatial and temporal dependencies, while foundation models demonstrate robust generalization across diverse climatic regions. The integration of these paradigms produces forecasts that are both more accurate and more interpretable. The findings highlight the potential of adaptive AI frameworks to improve environmental monitoring, detect regional anomalies, and strengthen climate resilience strategies in New Zealand.

Submission Number: 14