Flood generally refers to a hydrological phenomenon characterized by a rapid increase in water volume or water levels in rivers and lakes, or by excessive water accumulation on the land surface or in the soil. The flood disaster event records in CFDID are not equivalent to hydrological flood processes. Instead, CFDID adopts a source-based operational definition of flood disaster events. Because historical official documents do not consistently provide hydrological thresholds or fixed temporal windows for event delineation, event records in CFDID are delineated according to the event units used in the original official sources. Specifically, a flood-related disaster item was included when it was reported as a separate impact record in the source document and explicitly provided socioeconomic impacts, including casualties, affected population, affected cropland area, collapsed dwellings, direct economic losses, or other impact indicators. Records that describe only water level, streamflow, surface water accumulation, or soil moisture anomalies without socioeconomic impacts are not included.

To preserve the reporting structure of the original sources, CFDID uses a three-level event hierarchy. L1 records refer to parent-event records, representing independent flood disaster events or broad flood disaster events reported in the original sources. L2 records refer to sub-event records nested within L1 records, usually describing regional, local, or component impact records that belong to a broader L1 event. L3 records refer to more detailed nested records within L2 records. When the original sources reported both a broad disaster event and nested regional or local impact records, CFDID retained this hierarchy rather than redefining events using a uniform duration threshold. Therefore, L1, L2, and L3 should be interpreted as hierarchical record levels rather than three independent event types. For example, the L1 record “2011-043” describes the September 2011 flood impacts across six provincial-level regions. Its L2 child record “2011-043-01” reports impacts in Shaanxi. Nested within this L2 record, the L3 record “2011-043-01-01” reports localized impacts in Baqiao District, Xi'an.

Referring to the definitions in the 2024 Natural Disaster Statistical Investigation System issued by the Ministry of Emergency Management of China, CFDID classifies flood-related disasters into five categories according to the primary triggering mechanism: rainstorm flood, typhoon, snowmelt flood, ice-jam flood, and dam-break flood. These categories represent flood disasters mainly caused by heavy rainfall, typhoons, snowmelt, ice jams, and dam or dike failures, respectively. Outburst floods caused by the failure of barrier lakes are also classified as dam-break floods in CFDID. Because historical Chinese flood disaster records generally do not provide sufficient information to separate impacts among fluvial floods, pluvial floods, flash floods, and coastal floods, CFDID does not further classify events according to these common flood subtypes.

Several classification rules require further clarification. First, CFDID assigns each event record to a single primary disaster category according to the classification used in the original official sources. When multiple hazards occurred in the same disaster process, the category follows the main cause reported by the source document. Therefore, CFDID does not define a separate category for compound flood disasters. Second, the typhoon category should be interpreted as the event category reported in the original official sources, rather than as flood-only losses. Typhoon events may cause damage through strong winds, heavy rainfall, flooding, storm surges, and related secondary hazards. However, the source documents usually report the impacts of a typhoon event in aggregate and do not consistently separate the contributions of wind, rainfall, and flooding. Therefore, CFDID retains the reported aggregate impacts under the typhoon category. Third, CFDID does not systematically include all landslides, debris flows, or collapses. These geological hazards are included only when the original source explicitly states that they were triggered by rainstorms, floods, or typhoons and reports them as part of a flood- or typhoon-related disaster event. If the source does not provide a direct link to rainfall, flooding, or typhoon processes, these geological hazard impacts are not included in CFDID.

CFDID (1994–2023) is mainly compiled from 66 Chinese official documents. The core sources include the China Water Resources Yearbook, the Yearbook of Meteorological Disasters in China, and the annual list of the “Top Ten Natural Disasters in China” (Table 1). Examples of the three data sources and an event-description case are shown in Appendix A. These sources record the dates, locations, and main impact indicators of flood disasters in the corresponding years. The China Water Resources Yearbook and the Yearbook of Meteorological Disasters in China are mainly paper publications and were obtained through library access or purchase. The “Top Ten Natural Disasters in China” records are available from government websites. The China Water Resources Yearbook recorded event-scale flood impacts from 1994 to 2018, but systematic reporting of flood disaster information ceased after 2018 due to changes in government responsibilities. The Yearbook of Meteorological Disasters in China has provided relevant records since its first publication in 2002, while the “Top Ten Natural Disasters in China” list has been compiled since 2005. The two yearbooks document many flood events of varying magnitudes and the flood-related entries from the annual “Top Ten Natural Disasters in China” mainly list major-event information.

Another official source of flood-disaster statistics is the China Flood and Drought Bulletin, published by the Chinese water resources authorities. The flood-disaster statistics relevant to this study are the same as those reported in the China Water Resources Yearbook. Because the China Water Resources Yearbook has already been systematically incorporated into CFDID, we did not treat the Bulletin as an additional independent data source.

CFDID contains 22 fields, which can be grouped into event hierarchy and identifiers, temporal information, spatial information, disaster category, typhoon information, impact indicators, and data sources. The event level field records the hierarchical position of each record as L1, L2, or L3, while the event ID field provides a unique identifier that preserves this hierarchy. Temporal fields record the beginning and ending dates of direct disaster impacts. Spatial fields record affected provincial-level administrative regions and, when available, more specific place names reported in the original sources. The European flood-impact database HANZE (Paprotny et al., 2024) provided a methodological reference for the field design of CFDID, particularly the recording of event dates, locations, human and economic impacts, and data sources. CFDID's fields reflect the information available in Chinese official sources, including detailed agricultural and dwelling impact indicators.

Table 2 lists the field names, descriptions, data types, units, and requirement levels in the released database. The “Required” column indicates whether a field is mandatory, conditional, or optional. A “Yes” field must be recorded for each event record. The “conditional” typhoon field means that the typhoon ID or name should be recorded when the disaster category is typhoon. The “conditional” impact fields do not mean that every event must contain all impact indicators; rather, a record must contain at least one of the five core impact indicators: deaths and missing persons, affected population, affected cropland area, collapsed dwellings, or direct economic losses. Fields marked as “No” are optional and are recorded only when the original source provides the corresponding information. Missing values are recorded as null, indicating that the original source did not report the information; they should not be interpreted as zero.

CFDID retains direct economic losses at current prices as reported in the original sources, and provides losses adjusted to 2023 constant prices. The adjustment uses annual consumer price index (CPI), which is consistent with the treatment in EM-DAT and Wikimpacts (Delforge et al., 2025; Li et al., 2026). The CPI data used for this adjustment were obtained from the China Statistical Yearbooks for the corresponding years (NBS, 2025). For event i occurring in year t, the current-price loss is Impacti,tloss and the 2023 constant-price loss is calculated as: 1Impacti,2023loss=Impacti,tloss×CPI2023CPIt where CPIt and CPI2023 are the CPI indices for year t and 2023, respectively.

The construction workflow of CFDID consists of five steps: document preprocessing, rule-based segmentation, data extraction, data fusion, and data post-processing, as shown in Fig. 1.

The China Water Resources Yearbook (1994–2001) and the Yearbook of Meteorological Disasters in China (2002–2023) report annual national totals for five core flood impact indicators during 1994–2023. We calculated annual and total percentages relative to the corresponding official national totals to quantify the disaster impacts represented in CFDID and international disaster impact databases. Different databases use different event structures. For CFDID, only L1 records were used in the numerator, and nested L2 and L3 records were excluded to avoid double counting between parent events and their sub-events. For Wikimpacts, which also has a hierarchical structure, its L2 national-level records for China are used to select impacts, whereas its L1 records may span several countries. For EM-DAT and DFO, which do not have hierarchical event levels, the original event records were used directly. Percentages were calculated only for indicators available in each database. For database d, indicator k, and year t, the annual percentage relative to the corresponding official annual total is defined as: 2Ptd,k=∑i=1ntImpacti,td,kOfficialtk×100%

The total percentage relative to the corresponding official totals over 1994–2023 is calculated as: 3TPd,k=∑t=19942023∑i=1ntImpacti,td,k∑t=19942023Officialtk×100% where Impacti,td,k denotes the reported value of indicator k for event i in year t in database d. Officialtk denotes the corresponding official annual total of indicator k in year t. nt denotes the number of records from database d used in the analysis for year t.

The official annual totals used as benchmarks originate from the same governmental reporting system as the event-level records used to construct CFDID. Therefore, the comparison between CFDID and official annual totals is partly internal and should not be interpreted as a fully independent validation. Qiu et al. (2025) reported good consistency between Sentinel-1-derived flooded cropland areas and officially reported affected cropland areas in China during 2017–2021, which offers independent evidence for the affected-cropland indicator.

From 1994 to 2023, CFDID contains 1716 flood disaster event records across the Chinese mainland, including 1391 L1 event records, 300 L2 event records, and 25 L3 event records. Each event record contains start-year, start-month, and province information. Among the 1716 records, 1205 records contain deaths and missing persons; 1419 records contain affected population; 1241 records contain affected cropland area; 87 records contain damaged cropland area; 395 records contain failed cropland area; 969 records contain collapsed dwellings; 397 records contain damaged dwellings; and 1444 records contain direct economic losses.

Among the 1391 L1 event records, 142 contain nested L2 event records, and 22 of these L2 event records further contain nested L3 event records. The relatively small number of L2 and L3 records reflects the fact that the source documents used in this study are national-level government publications. These documents mainly report flood impacts at the provincial level, while sub-provincial details are recorded less systematically. Therefore, L2 and L3 records are mainly retained as supplementary information for detailed analyses of specific local events, whereas the following analyses are based on L1 event records unless otherwise stated.

The L1 records in CFDID represent substantial percentages of the official national totals for the five core impact indicators. Over 1994–2023, the corresponding total percentages are 68.49 % for deaths and missing persons, 72.24 % for affected population, 60.70 % for affected cropland area, 65.67 % for collapsed dwellings, and 74.86 % for direct economic losses (Fig. 2a). Comparisons with international disaster impact databases are provided in Sects. 3.6 and 3.7.

The annual percentages vary substantially among indicators and years (Fig. 2b–f). The annual percentage ranges from 35.70 % to 109.86 % for deaths and missing persons, from 28.83 % to 107.00 % for affected population, from 26.98 % to 98.41 % for affected cropland area, from 15.25 % to 105.19 % for collapsed dwellings, and from 51.22 % to 97.70 % for direct economic losses. Some annual percentages exceed 100 %, which may reflect differences in statistical approaches between event-scale records and official annual totals.

Across disaster categories, L1 event records in CFDID are dominated by rainstorm floods and typhoons (Fig. 3a). Among the 1391 L1 event records, 779 are rainstorm flood events, accounting for 56.00 %, and 601 are typhoon events, accounting for 43.21 %. Snowmelt floods, ice-jam floods, and dam-break floods are much less frequent, with 5, 2, and 4 events, respectively, together accounting for less than 1 % of all L1 event records.

Different disaster categories also contribute differently to disaster impacts. Rainstorm floods account for the largest shares of all five core impact indicators: 75.67 % of deaths and missing persons, 68.93 % of affected population, 70.70 % of affected cropland area, 77.89 % of collapsed dwellings, and 68.93 % of direct economic losses (Fig. 3b–f). Typhoon events rank second, contributing 23.88 % of deaths and missing persons, 31.03 % of affected population, 29.22 % of affected cropland area, 21.63 % of collapsed dwellings, and 30.69 % of direct economic losses. Snowmelt floods, ice-jam floods, and dam-break floods contribute only small shares to the five core impact indicators.

The annual number of L1 flood disaster event records in CFDID shows substantial interannual variability (Fig. 4a). During 1994–2023, the largest number of records occurred in 2008, with 106 events, followed by 2009 (102 events), 2007 (94 events) and 2005 (88 events). Overall, the number of recorded events was relatively high during the second decade, 2004–2013, with an average of 76.3 events per year. By contrast, the averages for the first decade, 1994–2003, and the third decade, 2014–2023, were lower, at 27.2 and 35.6 events per year, respectively. This pattern likely reflects differences in the level of detail across the source documents and should not be interpreted as direct evidence of long-term changes in hydrological flood frequency.

The annual distribution also differs among disaster categories (Fig. 4b–f). Rainstorm floods are the dominant category in most years, with the highest annual count of 80 events in 2009. Typhoon-related flood records vary considerably among years, with relatively high numbers in 2005, 2013, 2016, and 2018. Snowmelt floods, ice-jam floods, and dam-break floods are rare and occur intermittently.

Seasonally, L1 flood disaster event records are highly concentrated from May to October (Fig. 5a). During 1994–2023, 1335 events occurred during this period, accounting for 95.97 % of all L1 event records. July has the largest number of records, with 384 events, followed by August (381), June (222), September (181), May (94), and October (73). Seasonal patterns differ among disaster categories. Rainstorm floods mainly occur from May to August and peak in July, with 238 events. Typhoons mainly occur from June to October and peak in August, with 229 events. This seasonality is consistent with the concentrated summer rainfall in the East Asian monsoon region and the seasonal influence of western North Pacific typhoons on China (Shan et al., 2023).

Provincial statistics show clear spatial differences in L1 flood disaster event records (Fig. 6a). From 1994 to 2023, the 1391 L1 event records contain 1820 province-level records. Guangdong has the largest number of province-level records, with 155 records, followed by Guangxi (135), Sichuan (120), Fujian (115), Hunan (115), Jiangxi (113), Zhejiang (106), and Hubei (106). These high-count provinces are mainly located in the Pearl River Basin, the Yangtze River Basin, and the Southeast Rivers Basin, all of which are largely situated in southern China. The total number of province-level records is larger than the number of L1 event records because some L1 flood disaster events affected multiple provinces.

Different disaster categories show strong spatial heterogeneity. Rainstorm floods are mainly distributed in the Yangtze River Basin and the Southwest Rivers Basin. Sichuan has the highest number of rainstorm flood records, with 119 records, followed by Hubei (98), Hunan (95), and Guizhou (85) (Fig. 6b). Typhoon records are concentrated in the Pearl River Basin and the Southeast Rivers Basin, with the largest number of records in Guangdong (100), followed by Fujian (71), Guangxi (70), Hainan (67), and Zhejiang (63) (Fig. 6c). Snowmelt flood records are found only in Xinjiang, with five records (Fig. 6d). Ice-jam flood records are mainly distributed along the Yellow River, including Shaanxi and Inner Mongolia (Fig. 6e). Dam-break flood records are mainly found in western China, with three records in Xizang and one record each in Xinjiang, Yunnan, and Sichuan (Fig. 6f).

We compare CFDID with three international flood impact databases in terms of event record counts and reported impacts as percentages of official national totals (Table 4). To ensure consistency with the event hierarchy used in Wikimpacts, the comparison uses L1 records for CFDID and L2 national-level China records for Wikimpacts. For EM-DAT, we included all records classified as “Flood”, together with flood-related records corresponding to Landslide (wet), Rockfall (wet), Mudslide, Tropical cyclone, and Glacial lake outburst flood. For Wikimpacts, we used its L2 records associated with L1 main events classified as “Flood”, “Tropical Storm/Cyclone”, or “Extratropical Storm/Cyclone”. Database versions and access information are provided in Sect. 5.

During 1994–2023, CFDID contains 1391 L1 event records, exceeding the number of flood records for the Chinese mainland in EM-DAT, DFO, and Wikimpacts. However, record counts are influenced by both the information availability and differences in event delineation and aggregation, and should not be interpreted as the only measure of database completeness.

For the five core impact indicators, CFDID accounts for higher percentages of the corresponding official totals for affected population, collapsed dwellings, and direct economic losses than the international databases. In addition, CFDID provides agricultural impact indicators, including affected cropland area, damaged cropland area, and failed cropland area, which are rarely available in global databases. These indicators are particularly useful for evaluating agricultural flood impact models (Zhang et al., 2026).

The international databases each have their own strengths and can complement CFDID. For deaths and missing persons, EM-DAT accounts for a slightly higher percentage of the corresponding official total than CFDID. DFO provides spatial information derived from remote sensing and is valuable for locating high-impact flood events and characterizing affected or inundated areas. Wikimpacts provides a rich set of impact categories, although the percentages of the corresponding official totals represented in its China-related records remain relatively low at present. This may partly reflect the limited representation of Chinese-language flood-related news and reports in Wikipedia-based sources (Li et al., 2026).

Among the international databases considered, EM-DAT contains the largest number of event records and accounts for the highest percentages of the corresponding official totals for deaths and missing persons, affected population, and direct economic losses. We therefore further compare CFDID and EM-DAT with official annual totals for deaths and missing persons, affected population, and direct economic losses (Fig. 7a–c). Overall, the three series show broadly consistent interannual variability, indicating that both databases capture major fluctuations in flood impacts. The Pearson correlation coefficients between the CFDID and EM-DAT annual series during 1994–2023 are 0.933, 0.661, and 0.863 for deaths and missing persons, affected population, and direct economic losses, respectively. CFDID and EM-DAT show relatively similar annual patterns for deaths and missing persons, whereas CFDID reports higher affected population and direct economic losses than EM-DAT in most years.

Because annual totals can be dominated by large events, we further examined correspondence between CFDID L1 records and EM-DAT records using the matching procedure in Appendix C. The EM-DAT record “2018-0369-CHN” was excluded because its location information was insufficient, leaving 433 EM-DAT records for comparison with 1391 CFDID L1 records. In total, 313 EM-DAT records (72.29 %) have identified CFDID counterparts, involving 569 CFDID records (40.91 %). Figure 7d–f compares record counts across different magnitude classes, distinguishing records with and without identified counterparts. Missing values are displayed separately as N/A and are not treated as zero. Overall, 822 CFDID records (59.09 %) have no identified EM-DAT counterpart. The reported impacts of the unmatched CFDID records account for 14.43 %, 19.95 %, and 16.57 % of the official national totals for deaths and missing persons, affected population, and direct economic losses during 1994–2023, respectively.

Considering EM-DAT's two numerical entry criteria, 1246 CFDID L1 records report at least 10 deaths and missing persons or at least 100 affected persons. Of these, 540 have identified EM-DAT counterparts and 706 do not. The reported impacts of the latter records account for 13.67 %, 19.95 %, and 16.34 % of the official national totals for deaths and missing persons, affected population, and direct economic losses during 1994–2023, respectively. Conversely, 120 EM-DAT records have no identified CFDID counterpart. Their reported impacts represent 11.29 %, 2.23 %, and 5.09 % of the corresponding official national totals.

These results indicate that CFDID and EM-DAT provide complementary information on flood impacts. Further work could harmonize the spatial and temporal scopes of matched record groups to assess the consistency and complementary detail of their reported impacts. The identifiers of matched and unmatched records are provided with the CFDID release on 10.5281/zenodo.22918521 (Cui et al., 2026).

The main contribution of CFDID lies in transforming flood impact information scattered across official documents into an open, systematic, event-scale structured database, which has several potential applications.

First, CFDID can be used to assess changes in flood impacts in China, including trends in different impact indicators and the contributions of different drivers. Similar analyses have been conducted in Europe (Paprotny et al., 2018, 2025) and the United States (Davenport et al., 2021; Rashid et al., 2023). Second, CFDID can be used to validate flood impact assessment models. Although many studies have simulated or assessed flood impacts, their results often remain highly uncertain because of the lack of systematic validation data (Messori et al., 2026; Wing et al., 2020). Event-scale impact records can help improve model evaluation and support more reliable projections of future flood impacts in China. Third, CFDID can serve as an event inventory for investigating how extreme flood events lead to social and economic impacts, thereby supporting future flood prevention and risk reduction strategies (Garcia et al., 2025). CFDID can also be integrated with other global or regional datasets to provide a more reliable China component for future global studies of flood impacts.

Beyond scientific analysis, CFDID can also support practical risk management. For example, it can provide historical loss information for actuarial pricing and portfolio optimization by insurance and reinsurance companies, thereby contributing to the development of flood insurance and the improvement of insurance coverage in China (Cui et al., 2025).

Although CFDID includes more event records and provides higher percentages relative to official annual totals for most impact indicators than existing international databases, it still has several limitations.

First, CFDID currently relies mainly on central-government official records. Local government yearbooks, news reports, social media, academic papers, insurance company reports, international databases, and remote sensing products have not yet been systematically integrated. As a result, some small-scale flood disaster events may still be omitted, which is a common limitation of most natural disaster impact databases. Additional datasets could provide complementary information for future extensions of CFDID. Fu et al. (2025) constructed a county-level urban flood-event dataset for 2000–2022 from news reports, while Shen et al. (2025) identified urban flood events for 2012–2023 using Sina Weibo. Groundsource v1 also provides a global inventory of flood records extracted from news reports (Mayo et al., 2026). These news- and social-media-based inventories could help identify additional candidate events for future extensions of CFDID, although they do not systematically provide the full set of impact indicators included in CFDID. Remote-sensing products can provide complementary spatial information. For example, Qiu et al. (2025) mapped flooded croplands in China during 2017–2021 using Sentinel-1 observations at a much finer spatial resolution than administrative-unit-based disaster records. Physically derived hazard datasets, such as the Global Drought and Flood Catalogue (He et al., 2020), can also provide independent information on event timing and spatial extent.

Second, province-level spatial information has already been incorporated into Fig. 6. Each L1 event was assigned to all provinces listed in its “Province” field, and the resulting province-level event counts were mapped accordingly. However, for events affecting multiple provinces, the reported impacts cannot yet be fully disaggregated among individual provinces. Future work will therefore focus on finer-scale spatial disaggregation and on converting information recorded in the “Specific location” field, including prefectures, counties, and river basins, into spatial layers (Rosvold and Buhaug, 2021; Teber et al., 2026). This would facilitate more detailed spatial analyses and integration with hydrological, socioeconomic, and remote-sensing datasets.

Third, typhoon impacts in CFDID cannot be separated into wind and water components, and flood-related geological hazards cannot always be separated from flood impacts. In addition, CFDID classifies flood disasters according to triggering causes rather than common flood subtypes such as fluvial floods, pluvial floods, flash floods, and coastal floods. This limits subtype-level comparisons with databases from the United States, Japan, Europe, and other regions. The records of saltwater intrusion events along China's coasts during 2015–2024 compiled by Xu et al. (2026) could support the future expansion of coastal flood records in CFDID.

Fourth, CFDID is currently provided in tabular form. Future work will develop a visualization platform based on the spatially enriched database to present geographic information more effectively, add narratives of major flood disasters, and gradually build a knowledge base of flood disaster events in China.