AEFEcon-GPT: Fact-Grounded Language Generation from Multiscale Geospatial Foundation Representations

Authors:

Abstract
Geospatial foundation models encode geographic environments into compact representations, but their interpretation across spatial scales and time remains challenging. We investigate how frozen AlphaEarth Foundations (AEF) representations can generate descriptions of land-surface characteristics, urban functions, and socioeconomic conditions. We develop AEFEcon-GPT, which aligns multiscale AEF representations with a language model through a learned projection and two-stage semantic alignment. Structured labels from independent geographic datasets supervise regional types, attributes, and spatial–temporal relations; auxiliary attributes are not required during inference. The framework supports static description and spatial comparison at pixel, area-of-interest, and 1 km grid scales, and temporal interpretation at the grid scale. Experiments cover 36 Chinese cities during 2017–2024, with 167,088 supervision instances and 3,282 evaluation samples. Static interpretation achieves 0.620 exact-grade accuracy, 0.858 adjacent-grade accuracy, and 0.934 attribute coverage. Attribute-level accuracies are 0.608 for temporal change directions and 0.625 for spatial comparison relations. Representation diagnostics and ablations show that factual descriptions depend on both information encoded in AEF representations and structured semantic supervision. The framework enables verifiable geographic interpretation of geospatial foundation representations across scales and time.

This publication has 15 references indexed in Scilit: