Named Entity Recognition (NER) is a fundamental task in information extraction, yet general-purpose NER categories often fail to capture the specificity required for specialized domains such as climate change research. This paper presents a methodology for the automatic construction of a domain-specific NER type set with minimal supervision, leveraging a schema-based bottom-up approach to knowledge graph construction. The process begins with the identification of 655 core climate change-related terms, sourced from authoritative domain-specific resources. These terms are then semi-automatically aligned with Wikidata using SPARQL queries to take advantage of its hierarchical structure. A neighbourhood graph is constructed based on instance of (P31) and subclass of (P279) properties, forming the basis for community detection via the weighted Louvain algorithm. The resulting 59 communities are manually analyzed to derive a final set of 21 NER types, including Ecosystem, Energy Source, Natural Disaster, Meteorological Phenomenon, and Chemical. Validation against existing ontologies and terminological knowledge base (SWEET, ENVO, and EcoLexicon) reveals that the SWEET ontology provides the highest coverage, containing 57.25% of core terms and 65.38% of the proposed NER types. The findings demonstrate that integrating knowledge graphs, NLP-based information extraction, and community detection provides an effective approach for domain-specific NER schema construction.