OmniTerra: A Multi-Modal Spatio-Temporal Transformer Framework for Global Yield Intelligence and Carbon Sequestration Modeling

Implementing precision agriculture on the global scale requires combining satellite remote sensing, AI architectures such as deep learning, and the processing of live geospatial data streams. OmniTerra is a Multi-Modal Spatio-Temporal Transformer (ST-Transformer) framework for global crop yield intelligence and carbon sequestration modelling is presented. Using the Google Earth Engine (GEE) API, OmniTerra modules in real-time data, processing a feature comprising the Normalized Difference Vegetation Index (NDVI) into a transformer network with multi-head self-attention, which is a custom, end-toend learning model. The framework incorporates models and projections of vegetation dynamics for three dominant staplecrops in three agro-climatic zones: Wheat, Rice and Maize. The results show a confident inference of 94.2%, for deployment as an interactive web app using Streamlit Community Cloud. OmniTerra also introduces a carbon sequestration estimation module, based on the IPCC guidelines, which estimates ecosystem carbon storage at the field level. The evaluations based on a multi-country global yield dataset confirm the model's cross-geographical generalization, spanning North America, South America, and South Asia regions. This study fills the gap between frontier AI architectures and operational agricultural Precision Agricultureintelligence to take AI’s contribution to food security analytics to the next level.