MGIM: A Masked Modeling Framework for Land Parcel-Level Geo-Inference

About Masked Geographic Information Model(MGIM)Inspired by masked language models, this paper introduces the Masked Geographical Information Model (MGIM), a novel framework for learning contextual representations from spatio-temporal data. The framework's core innovations include a parcel-scale method for multi-source data fusion and a custom self-supervised masking strategy for diverse geographic elements. This integrated approach enables the model to effectively learn complex spatio-temporal relationships, leading to competitive performance across various reasoning tasks. Key results include: a position inference error of 504.15 m; crowd flow inference with a MAPE of 0.48 and RMSE of 30.4; urban event inference accuracy of 97.00%; a time inference error of 2.31 hours; and a parcel function similarity of 0.95. MGIM accurately reasons from spatio-temporal contexts and dynamically adjusts inferences according to contextual changes. Visualization of attention mechanisms further illustrates MGIM's capacity to construct contextually-aware representations and task-specific attention patterns analogous to natural language processing models. This study presents a new paradigm for general-purpose spatio-temporal modeling in real-world geographic scenarios, offering significant theoretical and practical value, and promising an effective solution for building a geographic foundation model.Please see “MGIM Project Document.pdf” for code implementation details. For more information on how to reproduce the results presented in the paper, refer to this document.