Geocoding converts unstructured geographic text into structured spatial data, which is crucial in fields such as urban planning, social media spatial analysis, and emergency response systems. Existing approaches predominantly model geocoding as a geographic grid classification task but struggle with the output space dimensionality explosion as the grid granularity increases. Furthermore, these methods generally overlook the inherent hierarchical structure of geographical texts and grids. In this paper, we propose a hierarchy-aware geocoding model based on cross-attention within the Seq2Seq framework, incorporating S2 geometry to model geocoding as a task for generating grid labels and predicting S2 tokens (labels of S2 grids) character-by-character. By incorporating a cross-attention mechanism into the decoder, the model dynamically perceives the address contexts at the hierarchical level that are most relevant to the current character prediction based on the input address text. Results show that the proposed model significantly outperforms previous approaches across multiple metrics, with a median and mean distance error of 41.46 m and 93.98 m, respectively. Furthermore, our method achieves superior results compared to others in regions with sparse data distribution, reducing the median and mean distance error by 16.27 m and 7.52 m, respectively, suggesting that our model has effectively mitigated the issue of insufficient learning in such regions.

Geocoding is a fundamental component of geographic information science that plays a crucial role in various geographical studies and applications involving text data. Current mainstream geocoding methods fall into two categories: geodesic-grid prediction and address matching. However, the geodesic-grid-prediction method’s localization accuracy is hindered by the density of grid partitioning, struggling to strike a balance between prediction accuracy and grid density. Address-matching methods mainly focus on the semantics of query text. However, they tend to ignore keyword information that can be used to distinguish candidates and introduce potential interference, which reduces matching accuracy. Inspired by the human map-usage process, we propose a two-stage address-matching approach that integrates geodesic-grid prediction and text-matching models. Initially, a multi-level text-classification model is used to generate a retrieval region proposal for an input query text. Subsequently, we search for the most relevant point of interest (POI) within the region-proposal area using a semantics-based text-retrieval model. We evaluated the proposed method using POI data from the Beijing Chaoyang District. The experimental results indicate that the proposed method provides high address-matching accuracy, increasing Recall@1 by 0.55 to 1.56 percentage points and MRR@5 by 0.54 to 1.68 percentage points.

Geospatial target localization refers to determining the geographic location of targets from natural language queries, which is essential for applications such as urban management, emergency response, and location-based services. As user queries evolve from address matching toward complex forms incorporating multiple spatial semantic factors, the localization process increasingly demands the coordination of heterogeneous tools and data sources. Automation remains hindered by the lack of a unified structured representation for semantic factors, the difficulty of translating complex spatial constraints into executable operations, and the unreliability of single-round planning and execution. This study proposes SGTAgent, an agent framework for geospatial target localization from complex spatial semantic queries. The parser transforms queries into structured spatial semantic factors, the planner organizes dependent spatial workflows, and the feedback mechanism distinguishes local parameter refinement from workflow restructuring. Experiments show that SGTAgent achieves higher localization accuracy than the evaluated generic baselines across multiple LLM backbones. Using DeepSeek-V4.1-Flash, SGTAgent achieves localization accuracies of 39.86%, 68.76%, and 81.88% at distance thresholds of 100, 300, and 500 m, respectively, outperforming the evaluated agents under the corresponding experimental configurations. On complex queries using GPT-4o, it achieves an LA@500m of 80.47%, compared with 70.12% for ReAct. Ablation experiments show that removing the complete Restate–Replan feedback mechanism reduces the 500 m localization accuracy from 81.88% to 74.83%, demonstrating the effectiveness of the execution feedback mechanism under the evaluated configuration. An independent evaluation of the parser achieves macro-F1 scores of 89.76% on the original benchmark and 89.62% on independently formulated queries, indicating consistent semantic factor extraction capability across different query formulations.