The performance of the proposed early warning method is evaluated against a traditional rainfall warning approach. The effectiveness of both methods is assessed using three standard metrics: hit, false alarm, and miss. A hit refers to a case where a village triggered a flash flood warning and a flash flood event actually occurred. A false alarm refers to a case where a warning was issued but no flash flood event occurred. A miss refers to a case where no warning was issued, yet a flash flood event occurred in the village (Wu, 2024).

Fujian Province is located along the southeast coast of China and lies within the mid-subtropical monsoon climate zone. The region is characterized by complex topography, where mountains and hills account for nearly 80% of the total land area. Influenced by the East Asian monsoon, rainfall in Fujian is highly uneven in both temporal and spatial distribution. The main flood season occurs from March to August, during which over 70% of the annual precipitation is concentrated. Rainfall events are often short in duration but high in intensity, and when combined with steep terrain, they make Fujian one of the provinces most prone to flash flood disasters in China.

To evaluate the proposed rainfall-based flash flood warning method that integrates spatial stratification and multi-mode discrimination, six representative villages in Fujian Province were selected as case studies. These villages are distributed across Longyan and Sanming cities, encompassing a wide range of catchment scales, topographic gradients, and rainfall station densities. Such selection enables comprehensive testing of the proposed method under diverse hydrological and geomorphological conditions. The chosen villages include both small catchments (e.g., Xipian Village and Xipu Village) and medium-to-large catchments (e.g., Xiaying Village and Guidongmen Community), ensuring the applicability and robustness of the method across different spatial scales.

The hydrological and geomorphological characteristics of the study villages and their upstream catchments are summarized in Table 1. The locations of the villages, their respective catchments, and the distribution of rainfall stations are shown in Figure 3. And the warning indicators for each village are summarized in Table 2.

For the evaluation of the proposed flash flood warning method, six recent representative rainfall events were selected to represent the dominant rainfall regimes and hydrological responses in Fujian across the six study villages. The rainfall processes at the associated warning stations and the corresponding disaster situations are summarized in Figure 4 and Table 3. They include both short-duration, high-intensity convective storms and longer-duration stratiform rainfall, as well as both flood-generating and non-flood-generating cases to evaluate hits and false alarms. Because the study villages span small to large catchments with varying gauge densities, the selected events also reflect differences in runoff concentration, spatial rainfall variability, and watershed sensitivity, ensuring a representative validation dataset.

To evaluate the effectiveness of the proposed method, flash flood warnings were simulated for six representative villages and compared with those generated by the traditional approach. The selection of the key thresholds was based on regional hydrometeorological characteristics and calibration experiments. The rain gauge density threshold y is influenced by climate, spatial rainfall variability, and the provincial mean station density, and typically ranges from 5 to 50 km²/station. The watershed area threshold A generally ranges from 50 to 200 km², with larger values suitable for steep basins and smaller values for flatter ones. Within these ranges, y = 15 km²/station and the A = 70 km² showed the best overall performance in our study. The multi-station proportion parameter β was calibrated through sensitivity tests, and β = 0.5 provided the most stable warning results.

As shown in Table 4, the proposed method achieved more accurate warnings for flood-affected villages while substantially reducing false alarms for non-affected villages. This demonstrates improved precision and reliability compared with the conventional single-station approach.

A sensitivity analysis was conducted to evaluate the impact of key parameters on the performance of the proposed warning method. Following the principle of controlling variables, four different parameter schemes were designed for comparative experiments, as outlined in Table 5. Scheme 1 represents the baseline configuration determined in Section 3.1.

1. Scheme 2: As y increased, the number of associated stations per village decreased (Table 6), retaining only Type 1 and Type 2 stations. Reduced coverage made the method more prone to false alarms during localized heavy rainfall events.

2. Scheme 3: When A was enlarged, some small watersheds (e.g., Xipu Village) were misclassified, causing single-station warnings to be triggered despite the absence of flooding (Figure 7). This indicates that an excessively high A can raise false alarm rates.

3. Scheme 4: With station associations unchanged, a lower β value reduced the threshold for triggering multi-station warnings, making the system more susceptible to false alarms, especially in basins with heterogeneous rainfall distribution.

While the proposed method demonstrates improved warning accuracy, its performance is subject to certain limitations. Sensitivity analysis indicates that the method’s effectiveness critically depends on proper parameter calibration, particularly the rainfall station density threshold ( y ) and the watershed area threshold ( A ). The framework assumes a reasonably dense rainfall gauge network; in data-sparse regions or areas with complex terrain, its ability to accurately capture rainfall spatial heterogeneity may be limited. Future work should focus on developing adaptive parameterization schemes to enhance applicability across diverse hydrological settings.