An Efficient Global Automatic Threshold Detection Algorithm for Large‐Scale Flood Distribution Analysis

Abstract Identifying the optimal threshold of a peaks over threshold (POT) series is crucial for effective flood distribution analysis and decision‐making for risk reduction. Here we propose a threshold detection method based on the Shuffled Complex Evolution (SCE‐UA) optimization algorithm that can automatically identify the global optimal threshold without any objective specification. Results show that the proposed method efficiently located the optimal threshold with fewer than approximately 4–13 times the number of goodness of fit tests and Anderson‐Darling tests compared to traditional methods at 10 river gauge stations across China. The automatically identified threshold matched well with the threshold identified by graphical diagnostics, and it reduced fitting biases of the generalized Pareto model over commonly used fixed thresholds. This detection method was subsequently applied to a large‐scale flood distribution analysis across 380 stations of the Eastern Monsoon Region of China. The range of optimal thresholds for the POT series was between 0.14 m 3 /s and 49,062.53 m 3 /s, with a median value of 293.55 m 3 /s for the 380 stations. The high‐flow threshold was particularly high in wet regions and low in arid/semiarid regions. It is also shown that small dry catchments with lower elevation, lower field capacity, and larger saturated hydraulic conductivity tend to display heavier flood tails (i.e., a higher probability of extreme flood occurrence). Our study demonstrates the potential of an SCE‐UA‐based threshold detection framework for large‐scale flood distribution analysis and also provides a general framework for automatic extraction of excess extremes.