Point-of-Interest (POI) recommendation has become increasingly important with the rise of location-based services. Graph Neural Networks (GNNs) have shown effectiveness in modeling high-order structural signals from user-POI interactions, but struggle to capture fine-grained spatio-temporal patterns and textual information. Meanwhile, Large Language Models (LLMs) offer powerful semantic reasoning capabilities, yet their integration into recommender systems is hindered by alignment gaps. To address these challenges, we propose a novel framework, LLM-Augmented Spatio-Temporal Graph Learning for POI recommendation (LSTGL), which integrates both structural and semantic information to improve the quality of POI recommendation. Specifically, LSTGL constructs a directed spatio-temporal interaction graph with edge weights encoding temporal and spatial proximity. A node-aware relational encoder propagates messages on this graph to learn contextual user and POI representations. Then it leverages a frozen LLM to infer user preferences and POI semantics via structured prompts based on review data. To bridge the modality gap, a dual-view Transformer fusion module is introduced to align and integrate structural and semantic representations. Extensive experiments on two real-world datasets demonstrate that LSTGL outperforms state-of-the-art baselines, achieving up to 4.08% relative gains in NDCG and 3.82% in Recall.