AutoGIS: an agent framework for automated geospatial data management and analysis

Geospatial analysis plays a critical role in urban governance, environmental monitoring, and disaster management, yet end-to-end automation remains constrained by fragmented data acquisition, complex operator coupling, and limited workflow reproducibility. We present AutoGIS, a contract-driven autonomous GIS agent framework that formulates geospatial analysis as a program synthesis problem governed by verifiable contracts and environmental feedback. AutoGIS comprises a data agent and a code agent. The data agent performs data-path-free discovery, acquisition, and processing through a Geospatial Asset Knowledge Graph (GAKG) and satisfaction gating, producing analysis-ready data and data contracts. The code agent conditions PyQGIS generation on these data contracts and algorithm contracts, and performs self-repair using stack traces and runtime logs. We further develop QGIS-GPT with a three-stage laddered alignment strategy to improve long-tail operator understanding and invocation. Experiments show that AutoGIS improves asset retrieval, task orchestration, algorithm understanding, code generation, and repair robustness across multi-source geospatial tasks. Case studies on buffer analysis, normalized difference vegetation index (NDVI) change analysis, and rule-based land-use classification demonstrate end-to-end intent-to-analysis capability, establishing a foundation for more reliable autonomous geospatial modeling.