Artificial Intelligence for Earth Observation and Next-Generation Earth System Modeling

Artificial intelligence (AI) is rapidly transforming both the acquisition of Earth observation (EO) data and the development of next-generation Earth system models (ESMs), addressing critical limitations in traditional data processing and physically based modeling frameworks. This study investigates how AI-driven methodologies, particularly deep learning, multimodal data fusion, and physics-informed approaches, enhance the extraction of geospatial information from large-scale EO datasets and enable more efficient, accurate, and scalable Earth system modeling. A systematic review methodology is adopted, synthesizing recent advances in AI applications for EO and hybrid AI–physics modeling frameworks. The analysis focuses on key architectures, including convolutional and transformer-based models for EO tasks, as well as physics-informed neural networks and hybrid ESMs that integrate process-based knowledge with data-driven learning. Emphasis is placed on evaluating their roles in environmental monitoring, disaster prediction, and climate modeling. The review shows that AI significantly improves performance in EO applications such as land-cover classification, change detection, and hazard monitoring while enabling the integration of heterogeneous datasets including optical, SAR, and LiDAR sources. Furthermore, hybrid and physics-informed AI models demonstrate enhanced predictive accuracy, computational efficiency, and generalization capabilities compared to conventional numerical models, particularly in simulating extreme events and complex climate processes. Emerging foundation models trained on large-scale geophysical datasets further illustrate the potential for unified, scalable Earth system intelligence. These findings suggest that the convergence of AI and Earth system science is shifting the paradigm from isolated data analysis and simulation toward integrated, learning-based modeling ecosystems. However, challenges remain in terms of model interpretability, data heterogeneity, physical consistency, and robustness under non-stationary climate conditions. Addressing these limitations will require standardized datasets, explainable AI frameworks, and closer integration between EO data streams and physics-based models. Future research should focus on developing transparent, hybrid modeling infrastructures and Earth digital twins that enable real-time, data-driven decision-making for climate resilience and sustainable environmental management.