Global Area Sampling for Geospatial Foundation Model

Abstract

In recent years, the field of geospatial analysis has seen significant advancements through the development of foundation models designed for specific regional applications. Last year, we introduced the Geospatial Foundation Model, called “Prithvi”, which demonstrated promising results in accurately capturing geospatial patterns and features through innovative area sampling methods. Building on this success, this presentation aims to extend the capabilities of the model to a global scale, addressing the complexities and challenges associated with diverse geographical regions and varied environmental conditions. We propose an extended area sampling framework that adapts to the heterogeneity of global landscapes, ensuring balanced and representative data collection. This framework leverages advanced machine learning algorithms to optimize sample selection, reducing bias and improving the overall accuracy of geospatial predictions. Through comprehensive experiments and case studies, we demonstrate the effectiveness of our global area sampling methods in capturing critical geospatial information. Our findings underscore the importance of scalable and adaptive sampling strategies in the development of robust geospatial foundation models.