Abstract
The rapid growth of sensor networks, satellite platforms, Internet-of-Things (IoT) devices, and urban monitoring systems has produced vast volumes of spatiotemporal data. Modeling such data is central to applications ranging from air-quality forecasting and climate monitoring to traffic prediction and urban planning, yet it poses challenges, including high dimensionality, spatial autocorrelation, and complex temporal dependencies that classical statistical models struggle to capture jointly. This review provides a comprehensive survey of spatiotemporal machine learning (STML), synthesizing the literature published between 2020 and 2025. We present a taxonomy of STML methods spanning traditional machine learning, recurrent neural networks (LSTM and GRU), convolutional architectures (3D CNN and ConvLSTM), graph neural networks, transformer-based models, and hybrid or ensemble designs. We organize representative studies by application domain, climate and environmental monitoring, transportation and traffic forecasting, and urban planning within smart cities, and provide a comparative analysis of models, datasets, evaluation metrics, and open-source tools. Finally, we identify emerging research trends and open challenges, including real-time processing, interpretability and explainability, scalability and edge deployment, multimodal fusion, federated learning for privacy, self-supervised learning, and ethical concerns such as fairness. The review is intended to serve as a structured reference for researchers and practitioners developing STML systems.