Comparative analysis of precipitation

Comparative Analysis of Precipitation

Bar : Precipitation in 10-minute intervals, black dotted line: Cumulative precipitation.

Graph A

AWS observation 10-minute interval precipitation time series

2022 Gangnam

Precipitation is highly concentrated within specific timeframes.
Precipitation was not significant initially, but it increased sharply after the evening of August 8. In particular, the cumulative precipitation rose rapidly as the 10-minute rainfall remained high for several hours.

= Typical characteristics of localized torrential rain

A case of intense precipitation concentrated within a short period, rather than rain falling gradually over a long time
In the graph, the black cumulative line also slopes sharply after the midpoint.

→ Clearly demonstrates the concentration of precipitation over a short period.

2023 Osong

The maximum rainfall per 10 minutes appears smaller than in 2022 Gangnam.
However, the rain continued repeatedly over several hours, and the cumulative rainfall steadily increased.
A case where damage was caused by continuous rainfall over a relatively long period, rather than a single moment of extreme precipitation.
Duration and cumulative rainfall are important, rather than looking only at the maximum rainfall intensity. → case of continuous rainfall accumulation.

2017 Cheongju

Precipitation was very distinctly concentrated within a single timeframe.
In particular, the bar grew rapidly during the morning hours of July 16, and the cumulative precipitation increased significantly within a short period thereafter.
It is similar to the 2022 Gangnam case: A case of localized torrential rain where strong precipitation was concentrated within a short period
Graphically, it shows a pattern where there is almost no rain before the precipitation begins, but the rainfall suddenly increases after a specific timeframe.
A pattern that is extremely difficult to forecast : Even if precipitation occurs only slightly late or in a different location, it may be classified as a Miss or False Alarm during forecast verification.

2011 Mt. Umyeon

The 2011 Mt. Umyeon case involved precipitation occurring over multiple periods.
Precipitation bars appeared continuously from the beginning to the middle of the event, and the cumulative line also steadily increased.
Rather than a case where precipitation was concentrated explosively at a single point in time like in Gangnam or Cheongju, it can be viewed as a case where repeated precipitation peaks amplified the risk of landslides.
Landslides cannot be explained solely by a single moment's maximum precipitation; they can be caused by the combined action of accumulated precipitation and ground saturation.

Visualization of precipitation over 10-minute intervals using AWS observation data revealed that localized torrential rain incidents were characterized by a rapid increase in precipitation over a short period or a rapid increase in cumulative precipitation as precipitation persisted for a certain duration. In particular, the 2022 Gangnam and the 2017 Cheongju case showed a distinct pattern of precipitation being concentrated in specific time periods, while the 2023 Osong and the 20211 Mt. Umyeon case demonstrate that an increase in cumulative precipitation can be linked to disaster damage.

Graph B

Comparison of cumulative precipitation by case

x-axis: time elapsed since the start of the event, y-axis: accumulated precipitation.

2022 Gangnam

The accumulation line rises very steeply in the middle section.
Precipitation accumulates explosively over a short period of time.

→ This can be directly linked to urban flooding. Urban drainage systems have a limit to the amount of precipitation they can handle per unit of time; as seen in the Gangnam case, if precipitation surges rapidly over a short period, it is easy to exceed drainage capacity.

2023 Osong

Cumulative precipitation increased steadily.
Rather than a pattern of sudden, rapid rise like in Gangnam, it is a form of accumulation resulting from rain continuing over a relatively long period.

→ A case where duration and cumulative precipitation were more important than maximum precipitation intensity.

2017 Cheongju

In the beginning, there was almost no increase, but after a certain point, the cumulative precipitation increased very rapidly.
Even a slight deviation in the forecast timing can result in a significant difference in performance evaluation.

2011 Mt. Umyeon

In 2011, Mt. Umyeon showed the highest or highest final cumulative precipitation.
However, since this case is linked to landslides, not only the simple amount of precipitation but also ground saturation due to cumulative precipitation must be considered.

A comparison of cumulative rainfall revealed that in all four cases, more than 250mm of rain accumulated during the event period. In particular, 2022 Gangnam and 2017 Cheongju showed a pattern of rapid increase in cumulative rainfall over a short period, while 2023 Osong and 2011 Mt. Umyeon are interpreted as cases where rainfall continued for a relatively long period, increasing the potential for damage.

Graph C

Maximum Hourly Precipitation
: Comparison of Observation, Forecast and Analysis

The graph compares the actual AWS observed maximum hourly precipitation and the KMA short-term forecast PCP maximum hourly precipitation for each case.

2022 Gangnam

The maximum PCP value in the short-term forecast was 14.0 mm/h, but the actual maximum observed by the AWS was 91.0 mm/h.

→ The actual precipitation intensity was about 6.5 times greater than the forecast.

2023 Osong

Forecast maximum: 30.0 mm/h
Observed maximum: 33.6 mm/h

→ The forecast predicted the precipitation intensity to be similar.

However, since this graph only compares maximum values, it does not necessarily mean that the prediction was successful.
If the time at which the forecast predicted 30 mm/h differs from the time at which the actual 33.6 mm/h was observed, it may be classified as Miss and False Alarm rather than Hit in the H/M/F verification.

2017 Cheongju & 2011 Mt. Umyeon

For both cases, observational data is available, but short-term forecast PCP data is unavailable. (No data)
Excluded from forecast-observation quantitative comparison because short-term forecast API data for the relevant historical time points was unavailable.

AWS Observations: Actual local precipitation extremes
Short-term forecast PCPs underestimate in some cases
ERA5 Analysis proxies appear much smaller in all cases

While ERA5 data offer the advantage of providing hourly data and covering wide periods, their spatial resolution is often described at approximately 30 km for ERA5 and 11 km for ERA5-Land, making it difficult to directly reproduce highly localized precipitation extremes occurring at AWS stations. In fact, limitations are also pointed out regarding ERA5 precipitation, as it may miss or underrepresent strong localized heavy rainfall.

Graph D

CSI Analysis Based on Actual Short-term Forecasts

The graph classifies Hit/Miss/False Alarms and calculating CSI by aligning forecasts and observations by hour.

Correct Negatives were excluded from the calculation.

2022 Gangnam

H = 0, M = 2, F = 0
In reality, there were two instances where heavy rainfall exceeding 30mm per hour occurred, but this implies that the short-term forecast failed to predict the heavy rain during those times.
In graph C as well, the forecast maximum value was 14.0mm/h, which is lower than the heavy rainfall threshold of 30mm/h, while the observed value was 91.0mm/h, which was very high.

→ C and D show the same direction.

The forecast significantly underestimated the precipitation intensity; as a result, a miss occurred during the hourly verification, where the actual heavy rainfall was missed.

2023 Osong

H = 0, M = 1, F = 1
In graph C, the forecast maximum was 30.0 mm/h and the observed maximum was 33.6 mm/h, which were similar, but in D, the CSI is 0.
There was one instance where the forecast indicated heavy rain, and one instance where actual heavy rain was observed, but the times did not coincide.
Therefore, the forecasted heavy rain time was classified as False Alarm, and the actual heavy rain time as Miss.
Although the precipitation intensity was somewhat similar, this is a case where it was not recognized as a Hit because the timing of occurrence was misaligned.

→ This demonstrates that for localized heavy rain, even a slight discrepancy in location and time can significantly lower prediction performance indicators.

2017 Cheongju & 2011 Mt. Umyeon

These two cases are excluded from the CSI calculation because although AWS observation data exists, short-term forecast PCP data is not available.

[D-1] Forecast CSI vs. Analysis CSI

We compared how accurately the Forecast and Analysis data matched the AWS observed heavy rainfall times.

→ The results showed that for both, the accuracy was mostly 0.

When the heavy rain criterion was set to 30mm or more per hour, neither Forecast nor Analysis produced a Hit. In particular, Analysis classified the actual heavy rain time as a Miss because the precipitation did not reach the heavy rain criterion in all cases.

[D-2] H/M/F Classification in Analysis

Hit = 0 - False Alarm = 0
Only Misses exist
While Analysis did not over-predict heavy rain at any time, it also did not detect the actual heavy rain. In other words, in these cases, Analysis demonstrated limitations centered on misses rather than false alarms.

Final Result

Analysis of AWS observation data revealed that all four heavy rainfall cases were characterized by a rapid increase in precipitation or a significant increase in cumulative precipitation over a short period.

When compared with actual short-term forecast PCPs, the 2022 Gangnam case was classified as a Miss because the forecast significantly underestimated the precipitation intensity, while the 2023 Osong case was not classified as a Hit because, although the maximum precipitation intensity was similar, the timing of occurrence was misaligned.

Therefore, it can be confirmed that in predicting localized heavy rainfall, the accuracy of the location and timing of occurrence, as well as precipitation intensity, is a key factor determining prediction performance.