Abstract: Objectives: Natural disaster events are characterized by fine-grained event types, complex semantic structures, and densely interconnected event relationships. These characteristics pose significant challenges for both event extraction and event relation identification tasks. Traditional classification-based approaches typically formulate event extraction as a multi-stage pipeline, where subsequent subtasks depend heavily on the outputs of earlier stages. In complex event scenarios, errors in event type identification are easily propagated and amplified across stages, leading to argument role mismatches and limiting the accurate modeling of event relations. In addition, argument extraction in these approaches often relies on local candidate-level predictions, which makes it difficult to accurately determine the semantic scope and boundaries of event arguments. Although generative models have recently demonstrated strong contextual understanding capabilities, they still suffer from issues such as event type hallucination and inaccurate argument role extraction and relation identification due to the absence of explicit structural constraints and domain knowledge. Methods: To address these challenges, this study proposes a systematic event ontology for natural disaster scenarios. The ontology categorizes events into three major classes: disaster events, result events, and response events, which are connected through three fundamental types of event relationships, namely causal, temporal, and concurrent relations. Based on this ontology, a natural disaster knowledge graph is constructed to provide structured domain knowledge for event understanding. Furthermore, we propose a knowledge graph-enhanced event extraction framework for natural disasters, named KGE³ (Knowledge Graph-Enhanced Event Extraction). In the proposed framework, a classification model is first combined with graph-guided matching to perform semantic rectification of event types, thereby improving event detection accuracy and reducing type ambiguity. Subsequently, natural language prompts are generated from the knowledge graph to guide a large language model in performing event argument extraction and event relation identification. To facilitate model training and evaluation in this domain, we also construct a dedicated dataset for natural disaster event extraction, named NDEE (Natural Disaster Event Extraction Dataset). Results: Extensive experiments conducted on the NDEE dataset demonstrate the effectiveness of the proposed method. KGE³ achieves F1 scores of 99.59%, 90.98%, and 81.51% for event detection, argument extraction, and event relation identification, respectively. Compared with existing baseline methods, the proposed approach shows consistent performance improvements across all tasks, with particularly notable gains of 3.67% in event detection and 3.74% in argument extraction. Conclusions: The experimental results verify that the proposed knowledge graph-enhanced framework effectively improves event type classification accuracy and structural consistency in event extraction. By integrating structured domain knowledge with large language models, KGE³ enables more accurate modeling of multi-event chains and effectively addresses the challenges posed by fine-grained type disambiguation and complex event structures in natural disaster scenarios.