Multimodal Climate Data Reconstruction in the Arabian Gulf Using Physics-Guided Artificial Intelligence

Introduction: Incomplete environmental observations remain a significant challenge for climate analysis in the Arabian Gulf, where measurement gaps may arise from limited station coverage, sensor failures, and adverse atmospheric conditions including dust storms and extreme weather events. Such deficiencies can introduce uncertainty into environmental assessments and climate applications. This study presents a physics-guided multimodal artificial intelligence framework for the reconstruction of missing climate observations and the enhancement of regional environmental datasets.

Methods: Atmospheric, hydrological, and land surface variables were integrated across multiple spatial and temporal scales to represent interactions among key Earth system processes. A transformer-based multimodal learning architecture was developed to capture relationships among temperature, precipitation, humidity, soil moisture, wind fields, and surface thermal characteristics. To improve physical consistency, constraints reflecting regional water and energy balance processes were incorporated into the model training procedure. Model performance was evaluated using systematic observation masking experiments in which a subset of available measurements was withheld and subsequently reconstructed.

Results: The proposed framework successfully reconstructed missing observations while preserving the spatial and temporal characteristics of the original datasets. Temperature reconstruction achieved a coefficient of determination exceeding 0.95, indicating strong agreement between reconstructed and reference observations. The multimodal architecture effectively represented coupled atmospheric and land surface processes and maintained stable performance during periods characterized by elevated temperatures and intense rainfall activity. The inclusion of physical constraints improved the consistency of reconstructed fields and reduced the occurrence of physically unrealistic patterns. Uncertainty analyses further demonstrated robust performance under varying levels of data availability.

Conclusions: The results demonstrate the capability of physics-guided artificial intelligence methods to improve the completeness and reliability of environmental datasets in regions affected by observational gaps. The proposed framework provides a basis for climate monitoring, environmental assessment, water resource management, and the future development of regional digital climate twin systems for the Arabian Gulf.