A human-verified database of reported socioeconomic drought impacts for 368 prefecture-level units in China, 2010–2025

Abstract

This database provides 5,888 unit-year records of reported socioeconomic drought impacts for 368 prefecture-level research units across China over the period 2010–2025. The research units comprise prefecture-level cities, county-level divisions directly administered by provincial-level governments, and the four municipalities directly under the central government (Beijing, Tianjin, Shanghai, and Chongqing). Owing to network access restrictions, Taiwan, Hong Kong, and Macao in China are not covered. Records were produced by a human-in-the-loop framework that couples a web search engine (Bocha AI Search) with a retrieval-augmented large language model (Qwen3-Max, Alibaba Cloud) to identify publicly reported socioeconomic drought impacts from Chinese web sources, followed by independent manual verification of every record against government bulletins, the Yearbook of Meteorological Disasters in China, and targeted web searches. A verified event is defined as any record confirmed by manual review, yielding 2,299 verified events. Each positive candidate is accompanied by a model-generated supporting statement quoting the evidence passages together with the source webpage links, enabling full traceability of every record. Confirmed events are further annotated with multi-label manifestation types covering supply-side impacts, demand-side impacts, emergency response, and other impacts. The database is stored as a single Excel workbook containing 17 worksheets: one Introduction sheet documenting the field definitions and coding scheme, and 16 annual sheets (2010–2025) with 368 rows each. Fields include Province (CN/EN), Research unit (CN/EN), Model Results (CN/EN), Supporting Webpage Links (CN), Manual Review Results (integer codes 0–4), and four multi-label Manifestation fields. Administrative fields are bilingual (Chinese and English); supporting statements and links are provided in Chinese only. Technical validation: manual verification of all 5,888 records gives an identification accuracy of 0.92, precision of 0.87, recall of 0.95, and F1 score of 0.91, with 112 confirmed LLM hallucination cases (1.9%); a 50-sample blind review of the manifestation labels gives an agreement rate of 93.0%. A companion Data Descriptor describing this database is under review at Scientific Data.