Geo-Mamba: A data-driven Mamba framework for spatiotemporal modeling with multi-source geographic factor integration

Abstract

Earth science data exhibit inherent complexity characterized by heterogeneous spatiotemporal attributes, high collinearity among variables, and diverse input formats. Despite rapid advancements in deep learning, geographic modeling lacks unified frameworks for integrating heterogeneous spatiotemporal data and diverse factor types. While Mamba architecture has demonstrated efficiency in large language models, computer vision and remote sensing, its applicability to geographic modeling remains unexplored. This study introduces Geo-Mamba, a novel framework addressing these challenges through three key innovations. First, we propose a systematic geographical factor classification method that categorizes elements into dynamic, static, and categorical factors, enabling standardized integration of heterogeneous data within a unified paradigm. Second, we design a selective encoder module based on Mamba architecture that leverages its linear complexity and scanning mechanism to establish selective state spaces for geographical inputs, revealing intricate associations between diverse feature types. Third, we incorporate Kolmogorov-Arnold Network (KAN) layers as intermediate components replacing multilayer perceptron linear layers, enhancing numerical regression accuracy in geographical applications. Experimental validation across three tasks demonstrates Geo-Mamba’s effectiveness: in net ecosystem exchange modeling (R2 = 0.92, RMSE = 0.37 μmol·m^–2·s^–1), groundwater storage anomaly downscaling (R² = 0.95, RMSE = 1.916 cm), and land cover classification (accuracy = 88.12 %, F1-Score = 84.27 %). These results confirm Geo-Mamba as an efficient unified framework for complex Earth science modeling, while establishing its viability for geographical data processing and factor integration.