Verification of Forecasted Three-Hour Accumulated 2 Precipitation Associated with “Senjo-Kousuitai” from Very- 3 Short-Range Forecasting Operated by the JMA 4

In recent years, “senjo-kousuitai”, characterized as band-shaped areas of heavy 40 rainfall, have frequently caused river floods and landslides in Japan. Preventing and 41 mitigating such disasters requires skillful forecasts of accumulated rainfall for several 42 hours with an adequate lead time. The immediate very-short-range forecast of 43 precipitation (VSRF) provided by the Japan Meteorological Agency (JMA) is well suited 44 to this purpose, representing a blended forecast of hourly accumulated precipitation for 45 up to 6 h ahead based on extrapolation and numerical weather prediction. This study 46 examined the predictability of the VSRF for 3-h accumulated precipitation associated with 47 21 senjo-kousuitai events that occurred in Kyushu in 2019 and 2020. Predictability was 48 evaluated based on forecast accuracy at each forecast time (1–6 h) using categorical and 49 neighborhood verification techniques. Overall, the VSRF product was useful for heavy 50 rainfall areas of ≥80 mm (3h) − 1 up to a forecast time of 2 h at the original grid spacing of 51 1 km, but with large uncertainty in the accuracy of the forecasts. After that forecast time, 52 it was not possible to obtain a useful precipitation forecast for the threshold of ≥80 mm 53 (3h) − 1 , even if displacement errors at municipal or larger scale (15–31 km) were tolerated. 54 Further analysis showed that the VSRF is less skillful in the stage of senjo-kousuitai 55 formation at shorter forecast times (1–2 h) owing to limitations of the extrapolation 56 forecasts. The poor skill during this period affects the timing of both issuance of warnings and decision-making regarding evacuation, representing major challenges for future development of forecasting methods and systems for senjo-kousuitai.


Introduction
In recent years, the occurrence of severe disasters in Japan caused by localized 65 and persistent heavy rainfall has increased (e.g., Danjo  as short-duration local thunderstorms, which implies likely increases in both the frequency 84 and the intensity of senjo-kousuitai in the future. Therefore, it is urgent that reliable methods 85 and systems be developed for forecasting senjo-kousuitai. 86 Because of the characteristic of persistent heavy rainfall of senjo-kousuitai, skillful

94
The JMA (2019) reported that in one heavy rainfall case, the immediate VSRF was able to 95 provide heavy rainfall information to the public approximately 20 min earlier for a 1-h forecast 96 because of the frequent and rapid updates of the forecasts. Therefore, the new VSRF could 97 be a product already in current operation that is suitable for predicting heavy rainfall 98 associated with senjo-kousuitai. Furthermore, this product has been used as input for 99 calculating the Soil Water Index that represents conceptual water stored in the soil (JMA 100 2019). In fact, considering this index, the JMA and the affected prefecture collaboratively 101 issue landslide alert information with a 2-h lead time to allow sufficient time for evacuation.

102
It indicates that the skill of VSRF could substantially affect the timing of both issuance of 103 5 warnings and decision-making regarding evacuation for sediment disasters associated with 104 senjo-kousuitai. 105 Given this background, the objective of this study was to quantitatively evaluate the 106 predictability of the VSRF for the 3-h accumulated precipitation (P3h) associated with senjo-  As forecast data, we used the immediate VSRF (hereafter, referred to simply as forecast is issued at 10-min intervals, which is higher frequency in comparison with the 122 conventional version (30-min update intervals). The VSRF employs a blending technique 123 6 that merges radar-based extrapolation with output from the JMA's operational NWP models 124 at an appropriate ratio. Generally, the extrapolation forecasts are more skillful than NWP 125 forecasts up to FTs of 1-2 h. However, their skill decreases rapidly with increasing FT 126 because the initiation, growth, and dissipation of precipitation systems are not considered.

127
On the other hand, the skill of NWP forecasts decreases gradually with increasing FT, which 128 can exceed that of extrapolation forecasts for longer FTs (e.g., Golding 1998; Sun et al.     FT. For all thresholds, the forecast accuracy is reasonably high at FT = 1 h, but it tends to 230 decrease rapidly within a few hours, especially for heavier precipitation (i.e., larger TH).

271
In this study, ∆t = 0 to 0.5 h, broadly corresponding to the formation stage of senjo-kousuitai. i.e., lower accuracy for shorter elapsed times. The FSS increases with increasing spatial 287 scale, but with less sensitivity at the formation stage (∆t = 0 h), which is probably attributable 288 to significant underestimation of the heavy precipitation area (Fig. 5a). Additionally, it is found forecasts are limited to around 10 min for meso-γ-scale localized heavy rainfall associated 310 with unorganized cumulonimbus clouds. A feature similar to that at FT = 2 h is also found at 311 FT = 1 h (Fig. S1a, b), where almost no NWP output is used for the forecast of 3-h 312 precipitation, indicating that forecast accuracy can be lower in the formation stage than the  (Fig. 2b), less precipitation might be observed at 317 the formation stage because it will correspond to the initiation of cumulonimbus development.

318
In contrast, the observations are expected to make a larger contribution after formation (e.g., 319 on and after ∆t = 1 h), which might lead to higher accuracy of the 3-h precipitation forecasts 320 in comparison with the formation stage. To exclude the effect of the observations, we also 321 conducted the same analysis as FT = 2 h for FT = 3 h (Fig. S1c, d), which consisted only of 322 forecast data (Fig. 2b). The result showed lower forecast accuracy at the formation stage,