Mobile QR Code QR CODE : Journal of the Korean Society of Civil Engineers

  1. ์ •ํšŒ์› ๐„ ์—ฐ์„ธ๋Œ€ํ•™๊ต ๊ฑด์„คํ™˜๊ฒฝ๊ณตํ•™๊ณผ ๋ฐ•์‚ฌํ›„ ์—ฐ๊ตฌ์›, ๊ณตํ•™๋ฐ•์‚ฌ (Yonsei University ๐„ tjghcjf1@gmail.com)
  2. ์ •ํšŒ์› ๐„ ์—ฐ์„ธ๋Œ€ํ•™๊ต ๊ฑด์„คํ™˜๊ฒฝ๊ณตํ•™๊ณผ ์„์‚ฌ๊ณผ์ • (Yonsei University ๐„ chul8456@yonsei.ac.kr)
  3. ์—ฐ์„ธ๋Œ€ํ•™๊ต ๊ฑด์„คํ™˜๊ฒฝ๊ณตํ•™๊ณผ ๋ฐ•์‚ฌํ›„ ์—ฐ๊ตฌ์›, ๊ณตํ•™๋ฐ•์‚ฌ (Yonsei University ๐„ suyeonc@yonsei.ac.kr)
  4. ์ •ํšŒ์› ๐„ ๊ต์‹ ์ €์ž ๐„ ์—ฐ์„ธ๋Œ€ํ•™๊ต ๊ฑด์„คํ™˜๊ฒฝ๊ณตํ•™๊ณผ ๊ต์ˆ˜, ๊ณตํ•™๋ฐ•์‚ฌ (Corresponding Author โ€ค Yonsei University ๐„ yeonjoo.kim@yonsei.ac.kr)



ConvLSTM, pySTEPS, ๊ฐ•์šฐ ์˜ˆ์ธก, ๋ ˆ์ด๋”
ConvLSTM, PySTEPS, Rainfall prediction, Radar

1. ์„œ ๋ก 

์ตœ๊ทผ ๊ธฐํ›„ ๋ณ€ํ™”์™€ ์ง€๊ตฌ ์˜จ๋‚œํ™”๋กœ ์ธํ•ด ๊ฐ•์šฐ ํŒจํ„ด์ด ๋ณ€ํ™”ํ•˜๊ณ , ๊ทน๋‹จ์ ์ธ ๊ธฐ์ƒ ํ˜„์ƒ์ด ๋นˆ๋ฒˆํ•˜๊ฒŒ ๋ฐœ์ƒํ•˜๊ณ  ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ณ€ํ™”๋Š” ๋„์‹œ ํ™์ˆ˜, ์ˆ˜์ž์› ๊ด€๋ฆฌ, ๋Œ ์šด์˜ ๋“ฑ ์—ฌ๋Ÿฌ ๋ถ„์•ผ์— ์‹ฌ๊ฐํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด์— ๋”ฐ๋ฅธ ํ”ผํ•ด๋ฅผ ์ตœ์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ •ํ™•ํ•œ ๋‹จ๊ธฐ ๊ฐ•์šฐ ์˜ˆ์ธก ๊ธฐ์ˆ ์˜ ์ค‘์š”์„ฑ์ด ๋”์šฑ ๋ถ€๊ฐ๋˜๊ณ  ์žˆ๋‹ค(Choi and Kim, 2022; Foresti et al., 2016; Giannone et al., 2008). ๋‹จ๊ธฐ ๊ฐ•์šฐ ์˜ˆ์ธก์€ ์ˆ˜ ์‹œ๊ฐ„ ๋‚ด์— ๋ฐœ์ƒํ•  ๊ฐ•์šฐ๋Ÿ‰์„ ์˜ˆ์ธกํ•˜๋Š” ๊ธฐ์ˆ ๋กœ, ์žฌ๋‚œ ์˜ˆ๋ฐฉ, ์‹ค์‹œ๊ฐ„ ํ™์ˆ˜ ๋Œ€์‘, ์ˆ˜์ž์› ๊ด€๋ฆฌ ๋ฐ ์šด์˜ ์ตœ์ ํ™” ๋“ฑ์— ํ•„์ˆ˜์ ์ธ ์—ญํ• ์„ ํ•œ๋‹ค(Carpenter and Georgakakos, 2004; Heuvelink et al., 2020).

๊ธฐ์กด์˜ ๋ ˆ์ด๋” ๊ธฐ๋ฐ˜ ๊ฐ•์šฐ ์˜ˆ์ธก ๋ชจ๋ธ์€ ์‹ค์‹œ๊ฐ„ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๊ฐ•์šฐ ํŒจํ„ด์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์œ ์šฉํ•˜๋‚˜, ๋ณต์žกํ•œ ๊ธฐ์ƒ ์กฐ๊ฑด์ด๋‚˜ ๋น„์„ ํ˜•์ ์ธ ๊ฐ•์šฐ ํŒจํ„ด์— ๋Œ€ํ•ด์„œ๋Š” ์˜ˆ์ธก ์ •ํ™•๋„์— ํ•œ๊ณ„๊ฐ€ ์กด์žฌํ•œ๋‹ค(Smith et al., 2024). ์ด๋Ÿฌํ•œ ํ•œ๊ณ„๋ฅผ ๋ณด์™„ํ•˜๊ณ ์ž Python framework for Short-Term Ensemble Prediction Systems(pySTEPS)์™€ Convolutional Long Short-Term Memory(ConvLSTM) ๊ธฐ๋ฐ˜์˜ ์ธ๊ณต์ง€๋Šฅ ๋ชจํ˜•์„ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์˜ˆ์ธก ์—ฐ๊ตฌ๊ฐ€ ํ™œ๋ฐœํžˆ ์ˆ˜ํ–‰๋˜๊ณ  ์žˆ๋‹ค.

pySTEPS๋Š” ํ†ต๊ณ„์  ๊ธฐ๋ฒ•๊ณผ ๊ด‘ํ•™ ํ๋ฆ„ ๊ธฐ๋ฐ˜์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ™œ์šฉํ•˜์—ฌ ๊ฐ•์šฐ ์˜ˆ์ธก์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ชจ๋ธ๋กœ, ์ƒ๋Œ€์ ์œผ๋กœ ๋†’์€ ๊ณ„์‚ฐ ํšจ์œจ์„ฑ๊ณผ ์•ˆ์ •์„ฑ์„ ์ œ๊ณตํ•œ๋‹ค(Pulkkinen et al., 2019). Smith et al.(2024)์€ pySTEPS๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ•ด์–‘ ๋Œ€๋ฅ™์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๊ฐ•ํ•œ ๋Œ€๋ฅ˜์„ฑ ํญํ’ ์˜ˆ์ธก ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ pySTEPS์˜ ๊ด‘ํ•™ ํ๋ฆ„ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์œ„์„ฑ ์˜์ƒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ๋ฆ„์˜ ์ด๋™์„ ์ถ”์ ํ•˜๊ณ  ๋ฏธ๋ž˜ ๊ฒฝ๋กœ๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ํšจ๊ณผ์ ์ธ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ, ๊ฒฐ์ •๋ก ์  ๊ธฐ๋ฒ•์€ ์•ฝ 4์‹œ๊ฐ„ ๋ฆฌ๋“œํƒ€์ž„์—์„œ ์ง€์†์„ฑ ์˜ˆ๋ณด๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ๋˜ํ•œ ํ™•๋ฅ ๋ก ์  ์˜ˆ์ธก ๋ฐฉ๋ฒ•์€ ์†Œ๊ทœ๋ชจ ๋Œ€๋ฅ˜์˜ ์ง€์†์‹œ๊ฐ„์— ๋Œ€ํ•œ ๋ถˆํ™•์‹ค์„ฑ์„ ์ค„์ด๋Š” ๋ฐ ๊ธฐ์—ฌํ•˜์˜€๊ณ , 60 km ์ดํ•˜์˜ ๊ณต๊ฐ„ ํ•ด์ƒ๋„์—์„œ ๋†’์€ ์˜ˆ์ธก ์ •ํ™•๋„๋ฅผ ๊ธฐ๋กํ•˜์˜€๋‹ค.

ConvLSTM์€ ์‹œ๊ณต๊ฐ„์  ์˜์กด์„ฑ์„ ๋™์‹œ์— ํ•™์Šตํ•˜์—ฌ ๋ณต์žกํ•œ ๊ฐ•์šฐ ํŒจํ„ด์„ ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜ ์˜ˆ์ธก ๋ชจํ˜•์ด๋‹ค (Shi et al., 2015). Kim et al.(2017)์€ ConvLSTM์„ ํ™œ์šฉํ•œ ๊ฐ•์šฐ ์˜ˆ์ธก ๋ชจ๋ธ์ธ DeepRain์„ ์ œ์•ˆํ•˜์˜€์œผ๋ฉฐ, 3์ฐจ์› ๊ธฐ์ƒ ๋ ˆ์ด๋” ๋ฐ์ดํ„ฐ๋ฅผ ์ž…๋ ฅ์œผ๋กœ ๋ฐ›์•„ ์‹œ๊ณ„์—ด ๊ฐ•์šฐ๋Ÿ‰์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐฉ์‹์„ ์ ์šฉํ•˜์˜€๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” 2๋…„๊ฐ„์˜ ๋ ˆ์ด๋” ๋ฐ˜์‚ฌ์œจ ์ž๋ฃŒ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์˜€๊ณ , 6๋ถ„ ๊ฐ„๊ฒฉ์œผ๋กœ ๊ตฌ์„ฑ๋œ ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ๋ฅผ ์ž…๋ ฅํ•˜์—ฌ ๊ฐ•์šฐ๋Ÿ‰์„ ์˜ˆ์ธกํ•˜์˜€๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, 2์ธต ๊ตฌ์กฐ์˜ ConvLSTM ๋ชจ๋ธ์€ ์„ ํ˜• ํšŒ๊ท€ ๋ชจ๋ธ ๋Œ€๋น„ RMSE๋ฅผ ์•ฝ 23 % ๊ฐ์†Œ์‹œํ‚ค๋ฉฐ, ConvLSTM ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ•์˜ ์šฐ์ˆ˜์„ฑ์„ ์ž…์ฆํ•˜์˜€๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋™์ผํ•œ ์•ˆ๋™๋Œ ์œ ์—ญ์„ ๋Œ€์ƒ์œผ๋กœ ๋ ˆ์ด๋” ์ž…๋ ฅ ๋„๋ฉ”์ธ ํฌ๊ธฐ(128 kmร—128 km, 256 kmร—256 km, 384 kmร— 384 km)์— ๋”ฐ๋ฅธ ConvLSTM๊ณผ pySTEPS์˜ ๊ฐ•์šฐ ์˜ˆ์ธก ์„ฑ๋Šฅ์„ ๋น„๊ตยท๋ถ„์„ํ•˜์˜€๋‹ค. 2014๋…„๋ถ€ํ„ฐ 2017๋…„๊นŒ์ง€์˜ ๋ ˆ์ด๋” ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šต์šฉ ์ž๋ฃŒ๋กœ ์‚ฌ์šฉํ•˜๊ณ , 2018๋…„ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฒ€์ฆ์šฉ์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ, ์ž…๋ ฅ ๋„๋ฉ”์ธ ํฌ๊ธฐ์— ๋”ฐ๋ฅธ ๋‘ ๋ชจ๋ธ์˜ ์˜ˆ์ธก ์„ฑ๋Šฅ ๋ณ€ํ™”๋ฅผ ๋ถ„์„ํ•˜์˜€๋‹ค.

2. ์—ฐ๊ตฌ ๋ฐฉ๋ฒ•

2.1 ์—ฐ๊ตฌ ์ง€์—ญ ๋ฐ ์ž…๋ ฅ์ž๋ฃŒ

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ ˆ์ด๋” ๊ฐ•์šฐ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์•ˆ๋™๋Œ ์œ ์—ญ์˜ ๊ฐ•์šฐ๋ฅผ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด, ์•ˆ๋™๋Œ ์ง€์—ญ์˜ ์ค‘์‹ฌ ์ขŒํ‘œ(128.77E, 36.85N)๋ฅผ ๊ธฐ์ค€์œผ๋กœ 128 kmร—128 km์˜ ์˜์—ญ์„ ์—ฐ๊ตฌ ์ง€์—ญ์œผ๋กœ ์„ค์ •ํ•˜์˜€๋‹ค(Fig. 1). ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์‚ฌ์šฉํ•œ ์ž…๋ ฅ์ž๋ฃŒ๋Š” ๊ธฐ์ƒ์ฒญ์—์„œ ์ œ๊ณตํ•˜๋Š” 1.5 km CAPPI ๋ ˆ์ด๋” ๋ฐ˜์‚ฌ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋ธ ํ•™์Šต ๋ฐ ํ‰๊ฐ€์— ํ™œ์šฉํ•˜์˜€๋‹ค. ํ•ด๋‹น ๋ฐ์ดํ„ฐ๋Š” ๋Œ€ํ•œ๋ฏผ๊ตญ ์ „์—ญ์˜ 11๊ฐœ ๋ ˆ์ด๋” ๊ด€์ธก์†Œ์—์„œ ์ˆ˜์ง‘๋œ ํ’ˆ์งˆ ๊ด€๋ฆฌ๋œ ํ•ฉ์„ฑ ๋ ˆ์ด๋” ๋ฐ˜์‚ฌ๋„(dBZ)๋กœ ๊ณต๊ฐ„ ํ•ด์ƒ๋„ 1 km, ์‹œ๊ฐ„ ํ•ด์ƒ๋„ 10๋ถ„, 960ร—1200 ํ”ฝ์…€์˜ ํ•ด์ƒ๋„๋ฅผ ๊ฐ€์ง„๋‹ค.

๊ฐ•์šฐ๋Ÿ‰์„ ์‚ฐ์ถœํ•˜๊ธฐ ์œ„ํ•ด ๋ ˆ์ด๋” ๋ฐ˜์‚ฌ๋„(dBZ)๋Š” Marshall- Palmer(1948)์˜ Z-R ๊ด€๊ณ„์‹(Eq. 1)์„ ํ™œ์šฉํ•˜์—ฌ ๊ฐ•์šฐ๊ฐ•๋„(mm/hr)๋กœ ๋ณ€ํ™˜๋˜์—ˆ๋‹ค.

(1)
$Z=200R^{1.6}$

์—ฌ๊ธฐ์„œ $Z$๋Š” ๋ ˆ์ด๋” ๋ฐ˜์‚ฌ๋„($mm^{6}/m^{3}$)๋ฅผ $R$์€ ๊ฐ•์šฐ๊ฐ•๋„($mm/hr$)๋ฅผ ์˜๋ฏธํ•œ๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์•ˆ๋™๋Œ ์œ ์—ญ์„ ์ค‘์‹ฌ์œผ๋กœ 128ร—128 km, 256ร— 256 km, 384ร—384 km ํฌ๊ธฐ์˜ ๋ ˆ์ด๋” ๊ฐ•์šฐ ์ถ”์ถœํ•˜์—ฌ 128ร—128, 256ร—256, 384ร—384 ํฌ๊ธฐ๋ฅผ ๊ฐ–๋Š” ๊ฐ•์šฐ ์ž๋ฃŒ๋ฅผ ์ƒ์„ฑํ•˜์˜€๋‹ค(Fig. 1). ์ƒ์„ฑ๋œ ๊ฐ•์šฐ ์ž๋ฃŒ๋Š” ๋™์ผํ•œ ํฌ๊ธฐ์ธ 128ร—128 km์˜ ์•ˆ๋™๋Œ ์œ ์—ญ์— ๋Œ€ํ•ด ๊ฐ•์šฐ ์˜ˆ์ธก ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์˜€๋‹ค. ๋˜ํ•œ, ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ชจ๋ธ๋“ค์˜ ์‹ค์‹œ๊ฐ„ ๊ฐ•์šฐ ์˜ˆ์ธก ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ 10๋ถ„์—์„œ 90๋ถ„์˜ ๋ฆฌ๋“œ ํƒ€์ž„(lead time)์„ ์„ค์ •ํ•˜์˜€๋‹ค. ๋ชจ๋ธ ํ•™์Šต์—๋Š” 2014๋…„๋ถ€ํ„ฐ 2017๋…„๊นŒ์ง€์˜ ๋ ˆ์ด๋” ๋ฐ์ดํ„ฐ๋ฅผ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋กœ ํ™œ์šฉํ•˜์˜€์œผ๋ฉฐ, 2018๋…„ ์—ฌ๋ฆ„์ฒ (6์›”~8์›”) ๋ฐ์ดํ„ฐ๋ฅผ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ํŠนํžˆ, ์—ฌ๋ฆ„์ฒ ์€ ๊ฐ•์šฐ๊ฐ•๋„๊ฐ€ ๋†’์•„ ๋‹ค์–‘ํ•œ ๊ธฐ์ƒ ์กฐ๊ฑด์„ ๋ฐ˜์˜ํ•œ ๊ฒฌ๊ณ ํ•œ ๋ชจ๋ธ ํ•™์Šต๊ณผ ๋…๋ฆฝ์ ์ธ ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ธฐ์— ์ ํ•ฉํ•œ ์‹œ๊ธฐ๋กœ ํŒ๋‹จ๋˜์—ˆ๋‹ค.

Fig. 1. Target Domain (the Andong Dam Basin) and Domain of Input Training Data

../../Resources/KSCE/Ksce.2025.45.3.0339/fig1.png

2.2 ConvLSTM

ConvLSTM์€ ํ•ฉ์„ฑ๊ณฑ ์—ฐ์‚ฐ๊ณผ LSTM(Long Short-Term Memory) ๊ตฌ์กฐ๋ฅผ ๊ฒฐํ•ฉํ•œ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ๋กœ, ์‹œ๊ฐ„์  ๋ฐ์ดํ„ฐ์™€ ๊ณต๊ฐ„์  ๋ฐ์ดํ„ฐ๋ฅผ ๋™์‹œ์— ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ•์ ์„ ์ง€๋‹Œ๋‹ค. ์ด๋Ÿฌํ•œ ํŠน์„ฑ์€ ๋น„๋””์˜ค ํ”„๋ ˆ์ž„, ์œ„์„ฑ ์ด๋ฏธ์ง€, ๊ธฐ์ƒ ๋ ˆ์ด๋” ๋ฐ์ดํ„ฐ ๋“ฑ ์‹œ๊ณต๊ฐ„ ์˜์กด์„ฑ์ด ๋†’์€ ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•˜๋Š” ๋ฐ ํŠนํžˆ ์ ํ•ฉํ•˜๋ฉฐ, ๊ฐ•์šฐ ์˜ˆ์ธก ๋ถ„์•ผ์—์„œ๋„ ํšจ๊ณผ์ ์œผ๋กœ ํ™œ์šฉ๋˜๊ณ  ์žˆ๋‹ค(Shi et al., 2015). ๊ธฐ์กด LSTM์ด 1์ฐจ์› ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ์— ์ตœ์ ํ™”๋œ ๋ฐ˜๋ฉด, ConvLSTM์€ ํ•ฉ์„ฑ๊ณฑ ์—ฐ์‚ฐ์„ ํ†ตํ•ด 2์ฐจ์› ๊ณต๊ฐ„ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์–ด ๋ณต์žกํ•œ ์ง€ํ˜•์  ํŠน์„ฑ์„ ๋ฐ˜์˜ํ•˜์—ฌ ๊ฐ•์šฐ ์˜ˆ์ธก ์ •ํ™•๋„๋ฅผ ๋†’์ด๋Š” ๋ฐ ์œ ๋ฆฌํ•˜๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ ๊ตฌํ˜„๋œ ConvLSTM ๋ชจ๋ธ์€ ๋ ˆ์ด๋” ๋ฐ˜์‚ฌ๋„ ์ด๋ฏธ์ง€ ์‹œํ€€์Šค์˜ ์‹œ๊ฐ„์  ๋ณ€ํ™”๋ฅผ ์ถ”์ ํ•˜๊ณ  ๊ณต๊ฐ„์  ํŠน์ง•์„ ๋™์‹œ์— ํ•™์Šตํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ๋‹ค. ๋ชจ๋ธ์€ ๋‹ค์ค‘ ConvLSTM2D ๋ ˆ์ด์–ด๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ๊ฐ ๋ ˆ์ด์–ด๋Š” 64๊ฐœ์˜ ํ•„ํ„ฐ์™€ 3ร—3 ์ปค๋„์„ ์‚ฌ์šฉํ•œ๋‹ค. ์ž…๋ ฅ ํฌ๊ธฐ๋ฅผ ์œ ์ง€ํ•˜๊ธฐ ์œ„ํ•ด โ€˜Sameโ€™ ํŒจ๋”ฉ์„ ์ ์šฉํ•˜์˜€๊ณ , ์•ˆ์ •์ ์ธ ํ•™์Šต ์ˆ˜๋ ด์„ ์œ„ํ•ด โ€˜HeNormalโ€™ ์ดˆ๊ธฐํ™” ๋ฐฉ๋ฒ•์„ ๋„์ž…ํ•˜์˜€๋‹ค.

๊ฐ ConvLSTM2D ๋ ˆ์ด์–ด ๋’ค์—๋Š” ๋ ˆ์ด์–ด ์ •๊ทœํ™”(Layer Normalization)๋ฅผ ์ ์šฉํ•˜์—ฌ ํ•™์Šต ์•ˆ์ •์„ฑ์„ ๊ฐ•ํ™”ํ•˜๊ณ , ๋‚ด๋ถ€ ๊ณต๋ณ€๋Ÿ‰ ๋ณ€ํ™”๋ฅผ ์ค„์˜€๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ๊ณผ ์ˆ˜๋ ด ์†๋„๋ฅผ ๊ฐœ์„ ํ•˜์˜€๋‹ค(Ba et al., 2016). ์ตœ์ข… ์ถœ๋ ฅ ๋ ˆ์ด์–ด๋Š” Conv3D ๋ ˆ์ด์–ด๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, (3, 3, 3) ์ปค๋„์„ ํ™œ์šฉํ•ด ์‹œ๊ฐ„์  ์ฐจ์›์— ๊ฑธ์นœ ๊ณต๊ฐ„์  ํŠน์ง•์„ ํ†ตํ•ฉํ•˜์—ฌ ๊ฐ•์šฐ๋Ÿ‰์„ ์˜ˆ์ธกํ•œ๋‹ค. ์ด๋Ÿฌํ•œ ๊ตฌ์กฐ๋Š” ๋ณต์žกํ•œ ๊ธฐ์ƒ ์กฐ๊ฑด์—์„œ๋„ ๋™์  ๊ฐ•์šฐ ์ด๋ฒคํŠธ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํฌ์ฐฉํ•˜๋Š” ๋ฐ ๋งค์šฐ ์œ ์šฉํ•˜๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ConvLSTM ๋ชจ๋ธ ํ•™์Šต์„ ์œ„ํ•ด ํ‰๊ท ์ œ๊ณฑ์˜ค์ฐจ(Mean sqaure error, MSE)๋ฅผ ์†์‹ค ํ•จ์ˆ˜๋กœ ์‚ฌ์šฉํ•˜์˜€์œผ๋ฉฐ, ์ตœ์ ํ™” ๊ธฐ๋ฒ•์œผ๋กœ ADAM ์˜ตํ‹ฐ๋งˆ์ด์ €๋ฅผ ์ ์šฉํ•˜์˜€๋‹ค. ํ•™์Šต๋ฅ ์€ 0.002๋กœ ์„ค์ •ํ•˜์˜€๊ณ , ์ด 400 ์—ํญ ๋™์•ˆ ํ•™์Šต์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค. ๋˜ํ•œ, ๊ณผ์ ํ•ฉ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์กฐ๊ธฐ ์ข…๋ฃŒ(Early Stopping) ๊ธฐ๋ฒ•์„ ๋„์ž…ํ•˜์˜€๊ณ , ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 32๋กœ ์„ค์ •ํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ ํšจ์œจ์„ฑ๊ณผ ๊ณ„์‚ฐ ์†๋„์˜ ๊ท ํ˜•์„ ์œ ์ง€ํ•˜์˜€๋‹ค.

2.3 pySTEPS

pySTEPS๋Š” ํ™•๋ฅ ์  ๊ฐ•์ˆ˜๋Ÿ‰ ๋‹จ๊ธฐ(0~6 h) ์˜ˆ์ธก์„ ์œ„ํ•ด ์„ค๊ณ„๋œ ์˜คํ”ˆ ์†Œ์Šค ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋กœ, ์—ฐ๊ตฌ์ž์™€ ๊ธฐ์ƒํ•™์ž๊ฐ€ ๋ ˆ์ด๋” ๊ธฐ๋ฐ˜ ๊ฐ•์ˆ˜ ์˜ˆ์ธก์„ ์ˆ˜ํ–‰ํ•˜๋ฉฐ ๋ถˆํ™•์‹ค์„ฑ์„ ์ •๋Ÿ‰ํ™” ํ•  ์ˆ˜ ์žˆ๋„๋ก ์œ ์—ฐํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ๊ณตํ•œ๋‹ค(Pulkkinen et al., 2019).

Smith et al.(2024)๋Š” pySTEPS์˜ ๊ด‘ํ•™ ํ๋ฆ„ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์œ„์„ฑ ๋ฐ์ดํ„ฐ์— ์ ์šฉํ•˜์—ฌ ๋Œ€๋ฅ˜์„ฑ ๊ฐ•์ˆ˜ ์˜ˆ์ธก์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜์˜€๋‹ค. ํŠนํžˆ, ํ•ด์–‘์„ฑ ๋Œ€๋ฅ™ ์ง€์—ญ์˜ ๋ณต์žกํ•œ ๊ฐ•์ˆ˜ ํŒจํ„ด์„ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ, pySTEPS๊ฐ€ ์œ„์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•œ ๊ฐ•์ˆ˜ ์˜ˆ์ธก์—์„œ๋„ ํšจ๊ณผ์ ์œผ๋กœ ์ ์šฉ๋  ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ์Œ์„ ํ™•์ธํ•˜์˜€๋‹ค. ์ด๋Š” ๋ ˆ์ด๋” ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๊ฑฐ๋‚˜ ๊ฐ€์šฉ์„ฑ์ด ์ œํ•œ๋œ ์ง€์—ญ์—์„œ๋„ pySTEPS๊ฐ€ ํ™œ์šฉ๋  ์ˆ˜ ์žˆ๋Š” ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ค€๋‹ค.

pySTEPS์˜ ํ•ต์‹ฌ์—๋Š” S-PROG(Spectral Prognosis) ๊ธฐ๋ฒ•์ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ†ต๊ณ„์  ์˜ˆ์ธก ๊ธฐ๋ฒ•์œผ๋กœ์„œ ๋ ˆ์ด๋” ๋ฐ์ดํ„ฐ์˜ ์‹œ๊ฐ„์  ๋ณ€ํ™”๋ฅผ ๋ถ„์„ํ•˜๊ณ  ํ”ฝ์…€ ์ด๋™์„ ์ถ”์ ํ•˜์—ฌ ๋ฏธ๋ž˜์˜ ๊ฐ•์ˆ˜๋Ÿ‰์„ ์˜ˆ์ธกํ•œ๋‹ค. ์†๋„์žฅ(velocity field)์„ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด PySTEPS๋Š” ๋ฃจ์นด์Šค-์นด๋‚˜๋ฐ ๋ฐฉ๋ฒ•(Lucas-Kanade method)์„ ํ™œ์šฉํ•œ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ๊ด‘ํ•™ ํ๋ฆ„ ๊ธฐ๋ฒ• ์ค‘ ํ•˜๋‚˜๋กœ, ์—ฐ์†๋œ ๋ ˆ์ด๋” ์ด๋ฏธ์ง€ ๊ฐ„์˜ ์†๋„ ๋ฒกํ„ฐ๋ฅผ ๊ณ„์‚ฐํ•œ๋‹ค. ์ด๋ ‡๊ฒŒ ์ƒ์„ฑ๋œ ์†๋„์žฅ์€ ์ดˆ๊ธฐ ๊ฐ•์ˆ˜ ํŒจํ„ด๊ณผ ๊ฒฐํ•ฉ๋˜์–ด S-PROG ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ์‹œ๊ฐ„ ๋‹จ๊ณ„๋ณ„๋กœ ์ง€์†์ ์ธ ๊ฐ•์ˆ˜ ์˜ˆ์ธก์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ๋‹ค.

์ถ”๊ฐ€์ ์œผ๋กœ, pySTEPS๋Š” ๊ฐ•์ˆ˜์˜ ์‹œ๊ณต๊ฐ„์  ๋ณ€ํ™”๋ฅผ ๋ณด๋‹ค ์ •๋ฐ€ํ•˜๊ฒŒ ๋ฐ˜์˜ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ํ™•๋ฅ ๋ก ์  ๋ฐฉ๋ฒ•์„ ์ ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ์„ค๊ณ„๋˜์—ˆ๋‹ค. ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์œ„์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ๋Œ€๋ฅ˜์„ฑ ๊ฐ•์ˆ˜ ์˜ˆ์ธก์˜ ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ํ•˜๋ฉฐ, pySTEPS๊ฐ€ ๊ฐ•์ˆ˜ ์‹œ์Šคํ…œ์˜ ์ด๋™ ๋ฐ ๋ณ€ํ™”๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋ชจ์‚ฌํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค.

2.4 ๋ชจ๋ธ ์„ฑ๋Šฅํ‰๊ฐ€

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ชจ๋ธ๋“ค์˜ ์˜ˆ์ธก ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ํ†ต๊ณ„ ์ง€ํ‘œ์™€ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์„ ์ ์šฉํ•˜์˜€๋‹ค. ์„ฑ๋Šฅ ํ‰๊ฐ€๋Š” ๋ฆฌ๋“œ ํƒ€์ž„์„ 10๋ถ„ ๊ฐ„๊ฒฉ์œผ๋กœ ์„ค์ •ํ•˜์—ฌ 10๋ถ„๋ถ€ํ„ฐ 90๋ถ„๊นŒ์ง€์˜ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋ฅผ ๋ถ„์„ํ•˜์˜€๋‹ค. ์ฃผ์š” ํ‰๊ฐ€์ง€ํ‘œ๋กœ๋Š” ์ƒ๊ด€๊ณ„์ˆ˜(R), RMSE, Nash-Sutcliffe ํšจ์œจ๊ณ„์ˆ˜(Nash-Sutcliffe Efficiency, NSE), ์ž„๊ณ„์„ฑ๊ณต์ง€์ˆ˜(Critical Success Index, CSI)๊ฐ€ ์‚ฌ์šฉ๋˜์—ˆ๋‹ค. ์ด๋“ค ์ง€ํ‘œ๋Š” ๊ฐ•์šฐ ์˜ˆ์ธก ์—ฐ๊ตฌ์—์„œ ๋„๋ฆฌ ํ™œ์šฉ๋˜๋Š” ์„ฑ๋Šฅ ํ‰๊ฐ€ ๊ธฐ์ค€์œผ๋กœ, ๋ชจ๋ธ์˜ ์˜ˆ์ธก ์ •ํ™•๋„์™€ ์‹ ๋ขฐ๋„๋ฅผ ์ •๋Ÿ‰์ ์œผ๋กœ ๋ถ„์„ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•œ๋‹ค.

์ƒ๊ด€๊ณ„์ˆ˜(R, Eq. 2)๋Š” ์˜ˆ์ธก๊ฐ’๊ณผ ๊ด€์ธก๊ฐ’ ์‚ฌ์ด์˜ ์„ ํ˜• ์ƒ๊ด€๊ด€๊ณ„์˜ ๊ฐ•๋„์™€ ๋ฐฉํ–ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ํ†ต๊ณ„์  ์ง€ํ‘œ์ด๋‹ค. R ๊ฐ’์ด 1์— ๊ฐ€๊นŒ์šธ์ˆ˜๋ก ๊ฐ•ํ•œ ์–‘์˜ ์„ ํ˜• ๊ด€๊ณ„๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ๊ด€์ธก๊ฐ’์ด ์ฆ๊ฐ€ํ•  ๋•Œ ์˜ˆ์ธก๊ฐ’๋„ ๋น„๋ก€์ ์œผ๋กœ ์ฆ๊ฐ€ํ•จ์„ ์˜๋ฏธํ•œ๋‹ค.

(2)
$R=\dfrac{\sum_{i=1}^{n}(P_{i}-\overline{P})(O_{i}-\overline{O})}{\sqrt{\sum_{i=1}^{n}(P_{i}-\overline{O})^{2}}\sqrt{\sum_{i=1}^{n}(O_{i}-\overline{O})^{2}}}$

์—ฌ๊ธฐ์„œ, $P_{i}$๋Š” ์˜ˆ์ธก๊ฐ’, $O_{i}$๋Š” ๊ด€์ธก๊ฐ’, $\overline{P}$๋Š” ์˜ˆ์ธก๊ฐ’์˜ ํ‰๊ท , $\overline{O}$๋Š” ๊ด€์ธก๊ฐ’์˜ ํ‰๊ท , $n$์€ ๊ด€์ธก๊ฐ’์˜ ์ด ๊ฐœ์ˆ˜๋ฅผ ์˜๋ฏธํ•œ๋‹ค.

RMSE (Eq. 3)๋Š” ์˜ˆ์ธก ์˜ค์ฐจ์˜ ํฌ๊ธฐ๋ฅผ ์ธก์ •ํ•˜๋Š” ๋Œ€ํ‘œ์ ์ธ ์„ฑ๋Šฅ ํ‰๊ฐ€ ์ง€ํ‘œ๋กœ, ์˜ˆ์ธก๊ฐ’๊ณผ ์‹ค์ œ ๊ด€์ธก๊ฐ’ ์‚ฌ์ด์˜ ์ œ๊ณฑ ํ‰๊ท  ์˜ค์ฐจ์˜ ์ œ๊ณฑ๊ทผ์œผ๋กœ ๊ณ„์‚ฐ๋œ๋‹ค.

(3)
$RMSE=\sqrt{\sum_{{i}=1}^{{n}}\dfrac{({P}_{{i}}-{O}_{{i}})^{2}}{{n}}}$

NSE (Eq. 4)๋Š” ๋ชจ๋ธ ์˜ˆ์ธก๊ฐ’๊ณผ ๊ด€์ธก๊ฐ’ ๊ฐ„์˜ ์ž”์ฐจ ๋ถ„์‚ฐ์„ ๊ด€์ธก ๋ฐ์ดํ„ฐ ๋ถ„์‚ฐ๊ณผ ๋น„๊ตํ•˜์—ฌ ๋ชจ๋ธ์˜ ์˜ˆ์ธก ์„ฑ๋Šฅ์„ ์ •๋Ÿ‰์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๋Š” ์ง€ํ‘œ์ด๋‹ค. NSE๋Š” -โˆž์—์„œ 1 ์‚ฌ์ด์˜ ๊ฐ’์„ ๊ฐ€์ง€๋ฉฐ, 1์€ ์™„๋ฒฝํ•œ ์˜ˆ์ธก ์„ฑ๋Šฅ์„ ์˜๋ฏธํ•œ๋‹ค. ๋ฐ˜๋ฉด, 0 ์ดํ•˜์˜ ๊ฐ’์€ ๋ชจ๋ธ์ด ๊ด€์ธก๊ฐ’์˜ ํ‰๊ท ์œผ๋กœ ์˜ˆ์ธกํ–ˆ์„ ๋•Œ์™€ ๊ฐ™๊ฑฐ๋‚˜ ๋” ๋‚ฎ์€ ์„ฑ๋Šฅ์„ ๋ณด์ž„์„ ์˜๋ฏธํ•œ๋‹ค.

(4)
$NSE=1-\dfrac{\sum_{i=1}^{n}(P_{i}-O_{i})^{2}}{\sum_{i=1}^{n}(O_{i}-\overline{O})^{2}}$

CSI (Eq. 5)๋Š” ๋ชจ๋ธ์ด ์˜ˆ์ธกํ•œ ์‚ฌ๊ฑด ์ค‘ ์‹ค์ œ ๊ด€์ธก๊ณผ ์ผ์น˜ํ•œ ์‚ฌ๊ฑด์˜ ๋น„์œจ์„ ํ‰๊ฐ€ํ•˜๋Š” ์„ฑ๋Šฅ ์ง€ํ‘œ์ด๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” 0.1 mm/h (CSI 0.1), 1.0 mm/h (CSI 1.0), 5.0 mm/h (CSI 5.0)์˜ ์„ธ ๊ฐ€์ง€ ๊ฐ•์šฐ ๊ฐ•๋„ ์ž„๊ณ„๊ฐ’์— ๋Œ€ํ•ด CSI๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ์ •๋Ÿ‰์ ์œผ๋กœ ๋ถ„์„ํ•˜์˜€๋‹ค.

(5)
$CSI=\dfrac{H}{H+F+M}$

์—ฌ๊ธฐ์„œ, $H$๋Š” ์ •ํ™•ํžˆ ์˜ˆ์ธกํ•œ ์‚ฌ๊ฑด, $F$๋Š” ๊ณผ๋Œ€์˜ˆ์ธกํ•œ ์‚ฌ๊ฑด, $M$์€ ๋ˆ„๋ฝ๋œ ์‚ฌ๊ฑด์„ ์˜๋ฏธํ•œ๋‹ค.

3. ๊ฒฐ ๊ณผ

3.1 ์ž…๋ ฅ ๋„๋ฉ”์ธ ํฌ๊ธฐ์— ๋”ฐ๋ฅธ ConvLSTM ๋ชจ๋ธ ์„ฑ๋Šฅํ‰๊ฐ€

Fig. 2๋Š” ์ž…๋ ฅ ๋„๋ฉ”์ธ ํฌ๊ธฐ์— ๋”ฐ๋ฅธ ๋ฆฌ๋“œํƒ€์ž„๋ณ„ ConvLSTM ๊ฐ•์šฐ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค.. ConvLSTM ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ ๋„๋ฉ”์ธ ํฌ๊ธฐ์™€ ๋ฆฌ๋“œ ํƒ€์ž„์— ๋”ฐ๋ผ ๋šœ๋ ทํ•œ ์ฐจ์ด๋ฅผ ๋ณด์˜€๋‹ค. ๋ฆฌ๋“œ ํƒ€์ž„์ด ์ฆ๊ฐ€ํ• ์ˆ˜๋ก ๋‹จ๊ธฐ ๊ฐ•์šฐ ์˜ˆ์ธก ์„ฑ๋Šฅ์ด ๊ธ‰๊ฒฉํžˆ ๊ฐ์†Œํ•˜๋Š” ๊ฒฝํ–ฅ์ด ํ™•์ธ๋˜์—ˆ๋‹ค. ํŠนํžˆ, ์ž…๋ ฅ ๋„๋ฉ”์ธ ํฌ๊ธฐ๊ฐ€ ๊ฐ€์žฅ ์ž‘์€ 128ร—128 ํฌ๊ธฐ์˜ ์ž๋ฃŒ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ•™์Šตํ–ˆ์„ ๋•Œ, ๋ฆฌ๋“œ ํƒ€์ž„์ด ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ๋” ๊ธ‰๊ฒฉํžˆ ์ €ํ•˜๋˜์—ˆ๋‹ค. ์ด๋Š” ์ž‘์€ ๋„๋ฉ”์ธ ํฌ๊ธฐ๊ฐ€ ์ œ๊ณตํ•˜๋Š” ๊ณต๊ฐ„์  ์ •๋ณด๊ฐ€ ์ œํ•œ์ ์ด์–ด์„œ, ๋ชจ๋ธ์ด ๊ฐ•์ˆ˜ ์‹œ์Šคํ…œ์˜ ์‹œ๊ฐ„์  ๋ฐ ๊ณต๊ฐ„์  ๋ณ€ํ™”๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ถ”์ ํ•˜๊ธฐ ์–ด๋ ค์›Œ์ง€๊ธฐ ๋•Œ๋ฌธ์œผ๋กœ ์ถ”์ธก๋œ๋‹ค.

128ร—128 ํฌ๊ธฐ์˜ ๊ฐ•์šฐ ๋ ˆ์ด๋” ๋ฐ์ดํ„ฐ๋ฅผ ์ด์šฉํ•œ ConvLSTM ๋ชจ๋ธ์˜ ํ•™์Šต ๊ฒฐ๊ณผ, ๋ฆฌ๋“œ ํƒ€์ž„ 10๋ถ„๊นŒ์ง€๋Š” ๊ฐ•์ˆ˜ ํŒจํ„ด์„ ๋น„๊ต์  ์ž˜ ํฌ์ฐฉํ•˜์˜€๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฆฌ๋“œ ํƒ€์ž„์ด 30๋ถ„์„ ์ดˆ๊ณผํ•˜๋ฉด์„œ ์˜ˆ์ธก ์„ฑ๋Šฅ์ด ๊ธ‰๊ฒฉํžˆ ๋–จ์–ด์กŒ์œผ๋ฉฐ, 60๋ถ„๊ณผ 80๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋Š” ๋Œ€๋ถ€๋ถ„์˜ ์ง€์—ญ์—์„œ ๊ฐ•์ˆ˜๋Ÿ‰์ด ๊ฑฐ์˜ 0์— ๊ฐ€๊นŒ์šด ๊ฐ’์œผ๋กœ ์˜ˆ์ธก๋˜์—ˆ๋‹ค. ๋ฐ˜๋ฉด, 256ร—256 ํฌ๊ธฐ์˜ ๋ ˆ์ด๋” ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ, ๋ชจ๋ธ์˜ ์˜ˆ์ธก ์„ฑ๋Šฅ์ด ๊ฐœ์„ ๋˜์—ˆ์œผ๋ฉฐ, ํŠนํžˆ 10๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋Š” ๋น„๊ต์  ๋ช…ํ™•ํ•œ ๊ฐ•์ˆ˜ ํŒจํ„ด์ด ์œ ์ง€๋˜์—ˆ๋‹ค. ํ•˜์ง€๋งŒ 30๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋Š” ์ผ๋ถ€ ์ •ํ™•๋„ ์ €ํ•˜๊ฐ€ ๋‚˜ํƒ€๋‚ฌ๊ณ , 60๋ถ„๊ณผ 80๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋Š” ์„ฑ๋Šฅ ์ €ํ•˜๊ฐ€ ์—ฌ์ „ํžˆ ๋‘๋“œ๋Ÿฌ์กŒ์ง€๋งŒ, 128ร—128 ๋ฐ์ดํ„ฐ ๋„๋ฉ”์ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค.

384ร—384 ํฌ๊ธฐ์˜ ๊ฐ•์šฐ ๋ ˆ์ด๋” ์ž๋ฃŒ๋ฅผ ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ, ์ „์ฒด์ ์œผ๋กœ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ConvLSTM ๋ชจ๋ธ์€ 10๋ถ„๊ณผ 30๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ ๋น„๊ต์  ์ •ํ™•ํ•œ ๊ฐ•์ˆ˜ ๊ตฌ์กฐ๋ฅผ ์œ ์ง€ํ–ˆ์œผ๋ฉฐ, 60๋ถ„๊ณผ 80๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋„ ์ผ๋ถ€ ๊ฐ•์ˆ˜ ํŒจํ„ด์„ ์œ ์ง€ํ•˜๋ฉด์„œ ๊ฐ•์šฐ ํŒจํ„ด์ด ํ๋ ค์ง€๋Š” ์˜ˆ์ธก์„ ๋‚˜ํƒ€๋ƒˆ๋‹ค.

Fig. 2. Results of Precipitation Predictions with the ConvLSTM Model for Different Domain Sizes (128ร—128, 256ร—256, 384ร—384) along the Lead Times

../../Resources/KSCE/Ksce.2025.45.3.0339/fig2.png

3.2 ์ž…๋ ฅ ๋„๋ฉ”์ธ ํฌ๊ธฐ์— ๋”ฐ๋ฅธ pySTEPS ๋ชจ๋ธ ์„ฑ๋Šฅํ‰๊ฐ€

Fig. 3์€ ์ž…๋ ฅ ๋„๋ฉ”์ธ ํฌ๊ธฐ์— ๋”ฐ๋ฅธ ๋ฆฌ๋“œํƒ€์ž„๋ณ„ pySTEPS ๊ฐ•์šฐ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. pySTEPS ๋ชจ๋ธ๋„ ConvLSTM ๋ชจ๋ธ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ž…๋ ฅ๋„๋ฉ”์ธ์˜ ํฌ๊ธฐ๊ฐ€ ์ปค์งˆ์ˆ˜๋ก ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ๊ฐœ์„ ๋˜๋Š” ๊ฒฝํ–ฅ์„ ๋ณด์ธ๋‹ค(Fig. 3). 128ร—128 ํฌ๊ธฐ์˜ ์ž…๋ ฅ ๋„๋ฉ”์ธ์œผ๋กœ pySTEPS ๋ชจ๋ธ์„ ํ•™์Šตํ•œ ๊ฒฐ๊ณผ, 10๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋Š” ๊ฐ•์ˆ˜ ํŒจํ„ด์„ ์ผ๋ถ€ ํฌ์ฐฉํ•  ์ˆ˜ ์žˆ์—ˆ์œผ๋‚˜, ์˜ˆ์ธก ์ •ํ™•๋„๋Š” ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์•˜๋‹ค. ๋ฆฌ๋“œ ํƒ€์ž„์ด ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋ชจ๋ธ์˜ ์˜ˆ์ธก ์„ฑ๋Šฅ์€ ๊ธ‰๊ฒฉํžˆ ์ €ํ•˜๋˜์—ˆ์œผ๋ฉฐ, 30๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋Š” ๋Œ€๋ถ€๋ถ„์˜ ์˜ˆ์ธก ๊ฐ•์ˆ˜ ํŒจํ„ด์ด ๋‹จ์ˆœํ™”๋˜์—ˆ๊ณ , 60๋ถ„ ๋ฐ 80๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋Š” ์˜ˆ์ธก๋œ ๊ฐ•์ˆ˜๋Ÿ‰์ด ๊ฑฐ์˜ 0์— ๊ฐ€๊นŒ์šด ๊ฐ’์œผ๋กœ ์ˆ˜๋ ดํ–ˆ๋‹ค.

256ร—256 ํฌ๊ธฐ์˜ ๊ฐ•์šฐ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ, pySTEPS์˜ ์„ฑ๋Šฅ์ด ๊ฐœ์„ ๋˜์—ˆ์œผ๋ฉฐ, 10๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋Š” ๊ฐ•์ˆ˜ ๊ตฌ์กฐ๊ฐ€ ๋น„๊ต์  ๋ช…ํ™•ํ•˜๊ฒŒ ์œ ์ง€๋˜์—ˆ๋‹ค. 30๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋„ ์˜ˆ์ธก๋œ ๊ฐ•์ˆ˜ ํŒจํ„ด์€ ์ƒ๋Œ€์ ์œผ๋กœ ์ž˜ ๊ตฌ๋ถ„๋˜์—ˆ์ง€๋งŒ, 60๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„ ์ดํ›„์—๋Š” ์„ฑ๋Šฅ ์ €ํ•˜๊ฐ€ ๋‚˜ํƒ€๋‚ฌ๊ณ , ์˜ˆ์ธก๋œ ๊ฐ•์ˆ˜๋Ÿ‰์€ ์ดˆ๊ธฐ ๋ฆฌ๋“œ ํƒ€์ž„์— ๋น„ํ•ด ํ˜„์ €ํžˆ ์•ฝํ•ด์กŒ๋‹ค.

ConvLSTM ๋ชจ๋ธ๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ, pySTEPS ๋ชจ๋ธ๋„ 384ร—384 ํฌ๊ธฐ์˜ ์ž…๋ ฅ ๋„๋ฉ”์ธ์„ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. 30๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„๊นŒ์ง€๋Š” ๊ฐ•์ˆ˜ ํŒจํ„ด์ด ๋น„๊ต์  ๋ช…ํ™•ํžˆ ์œ ์ง€๋˜์—ˆ์œผ๋ฉฐ, 60๋ถ„ ๋ฐ 80๋ถ„ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ๋„ ๋‹ค๋ฅธ ๋ชจ๋ธ์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚ฌ๋‹ค. ์šฐ๋ฆฌ๋‚˜๋ผ ์—ฌ๋ฆ„์ฒ  ๊ฐ•์ˆ˜๋Ÿ‰์˜ ์ด๋™ ํŒจํ„ด์€ ์ฃผ๋กœ ๋‚จ์—์„œ ๋ถ์œผ๋กœ ์ด๋™ํ•˜๋Š” ๊ฒฝํ–ฅ์ด ๋‚˜ํƒ€๋‚œ๋‹ค. ๋™์ผํ•œ ์ง€์—ญ๊ณผ ์ž…๋ ฅ ๋„๋ฉ”์ธ์„ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ, ๋ฆฌ๋“œ ํƒ€์ž„์ด ๊ธธ์–ด์งˆ์ˆ˜๋ก ๋‚จ์ชฝ ์ง€์—ญ์˜ ๊ฐ•์ˆ˜๋Ÿ‰์„ ์˜ˆ์ธกํ•˜์ง€ ๋ชปํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์—ˆ์ง€๋งŒ, ํ•ด๋‹น ์œ ์—ญ๋ณด๋‹ค ๋” ํฐ ์ž…๋ ฅ ๋„๋ฉ”์ธ์œผ๋กœ ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๋ฉด ๋ฆฌ๋“œ ํƒ€์ž„์ด ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋‚จ์ชฝ ์ง€์—ญ ๊ฐ•์ˆ˜๋Ÿ‰์„ ๋” ์ž˜ ์˜ˆ์ธกํ•˜๋Š” ๊ฒฝํ–ฅ์„ ๋ณด์˜€๋‹ค.

Fig. 3. Results of Precipitation Predictions with the pySTEPS Model for Different Domain Sizes (128ร—128, 256ร—256, 384ร—384) along the Lead Times

../../Resources/KSCE/Ksce.2025.45.3.0339/fig3.png

3.3 ๊ฐ•์šฐ ์˜ˆ์ธก ์„ฑ๋Šฅ ๋น„๊ต ๋ถ„์„

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ConvLSTM๊ณผ pySTEPS ๋‘ ๋ชจ๋ธ์˜ ๊ฐ•์šฐ ์˜ˆ์ธก ์„ฑ๋Šฅ์„ ์ž…๋ ฅ ๋„๋ฉ”์ธ ํฌ๊ธฐ(128ร—128, 256ร—256, 384ร—384)์™€ ๋ฆฌ๋“œ ํƒ€์ž„(10๋ถ„, 30๋ถ„, 60๋ถ„, 90๋ถ„)์— ๋”ฐ๋ผ ๋น„๊ต ๋ถ„์„ํ•˜์˜€๋‹ค. ๋„๋ฉ”์ธ ํฌ๊ธฐ ์ฆ๊ฐ€์— ๋”ฐ๋ฅธ ์„ฑ๋Šฅ ๊ฐœ์„  ๋น„์œจ์€ ์ƒ๊ด€๊ณ„์ˆ˜, RMSE, NSE, CSI(0.1 mm, 1.0 mm, 5.0 mm ๊ฐ•์šฐ๊ฐ•๋„)๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์ •๋Ÿ‰์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜์˜€์œผ๋ฉฐ, Fig. 4์™€ Table 1์—์„œ ์‹œ๊ฐ์  ๋ฐ ์ •๋Ÿ‰์ ์œผ๋กœ ์ œ์‹œํ•˜์˜€๋‹ค.

Fig. 4๋Š” ์ƒ๊ด€๊ณ„์ˆ˜, RMSE, NSE ๊ฒฐ๊ณผ๋ฅผ ํ†ตํ•ด ๋„๋ฉ”์ธ ํฌ๊ธฐ์™€ ๋ฆฌ๋“œ ํƒ€์ž„์— ๋”ฐ๋ฅธ ๋‘ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ๋ณ€ํ™”๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ConvLSTM์€ ๋ฆฌ๋“œ ํƒ€์ž„์ด ์ฆ๊ฐ€ํ• ์ˆ˜๋ก ์„ฑ๋Šฅ ์ €ํ•˜๊ฐ€ ๋‘๋“œ๋Ÿฌ์กŒ์œผ๋ฉฐ, ํŠนํžˆ 128ร—128 ํฌ๊ธฐ์˜ ์ž…๋ ฅ ๋„๋ฉ”์ธ์„ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ ์ƒ๊ด€๊ณ„์ˆ˜๊ฐ€ ๊ธ‰๊ฒฉํžˆ ๊ฐ์†Œํ•˜์˜€๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ 384ร—384 ํฌ๊ธฐ์˜ ์ž…๋ ฅ ๋„๋ฉ”์ธ ์ž๋ฃŒ๋ฅผ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ๋Š” ๋†’์€ ์ƒ๊ด€๊ณ„์ˆ˜๋ฅผ ์œ ์ง€ํ•˜๋ฉฐ ์ƒ๋Œ€์ ์œผ๋กœ ์•ˆ์ •์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ๋˜ํ•œ, Fig. 4์—์„œ๋Š” CSI ์ง€ํ‘œ(0.1 mm, 1.0 mm, 5.0 mm ๊ฐ•์šฐ ๊ฐ•๋„)์— ๋”ฐ๋ฅธ ์„ฑ๋Šฅ ๋ณ€ํ™”๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ๋‚ฎ์€ ๊ฐ•์šฐ ๊ฐ•๋„(0.1 mm)์—์„œ๋Š” ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ ๋†’์€ CSI ๊ฐ’์„ ๊ธฐ๋กํ–ˆ์œผ๋‚˜, ๊ฐ•์šฐ ๊ฐ•๋„๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ ์ €ํ•˜๊ฐ€ ๋‚˜ํƒ€๋‚ฌ๋‹ค. ํŠนํžˆ ConvLSTM์€ ๋†’์€ ๊ฐ•์šฐ ๊ฐ•๋„(5.0 mm ์ด์ƒ)์—์„œ ์ •ํ™•๋„๊ฐ€ ํฌ๊ฒŒ ๊ฐ์†Œํ–ˆ์ง€๋งŒ, pySTEPS๋Š” ์ƒ๋Œ€์ ์œผ๋กœ ์•ˆ์ •์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์ด๋ฉฐ ๋†’์€ CSI ๊ฐ’์„ ์œ ์ง€ํ•˜์˜€๋‹ค.

pySTEPS ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ๋„ ConvLSTM๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๋ฆฌ๋“œ ํƒ€์ž„์ด ์ฆ๊ฐ€ํ• ์ˆ˜๋ก ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ค„์–ด๋“ค๊ณ  ์ž…๋ ฅ๋„๋ฉ”์ธ์˜ ํฌ๊ธฐ๊ฐ€ ํด์ˆ˜๋ก ๋ชจ๋ธ์— ์„ฑ๋Šฅ์ด ์ฆ๊ฐ€ํ•œ๋‹ค. ํ•˜์ง€๋งŒ ๋ฆฌ๋“œ ํƒ€์ž„์ด 90๋ถ„์ผ ๋•Œ, ๋Œ€์ฒด๋กœ pySTEPS์˜ ์„ฑ๋Šฅ์ด ConvLSTM๋ณด๋‹ค ๋” ์šฐ์ˆ˜ํ•จ์„ ํ™•์ธํ•˜์˜€๋‹ค. ์ด๋Š” ์•ฝ 90๋ถ„ ์ •๋„์˜ ๋‹จ๊ธฐ ๊ฐ•์šฐ ์˜ˆ์ธก์—์„œ๋Š” 2๋ฐฐ ๋” ๋„“์€ ์ž…๋ ฅ ๋„๋ฉ”์ธ๊ณผ pySTEPS ๋ชจ๋ธ์„ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ ์„ฑ๋Šฅ์ด ๋” ์šฐ์ˆ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•œ๋‹ค.

Table 1์€ ConvLSTM๊ณผ pySTEPS ๋ชจ๋ธ์˜ 128ร—128 ๋„๋ฉ”์ธ์„ ๊ธฐ์ค€์œผ๋กœ, 256ร—256 ๋ฐ 384ร—384 ๋„๋ฉ”์ธ์—์„œ์˜ ์„ฑ๋Šฅ ๊ฐœ์„  ๋น„์œจ์„ ์ˆ˜์น˜ํ™”ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ConvLSTM์€ 384ร—384 ๋„๋ฉ”์ธ์—์„œ ์ƒ๊ด€๊ณ„์ˆ˜ 10.4 %, RMSE 10.1 %, NSE 10.3 %, CSI 0.1์—์„œ ์ตœ๋Œ€ 15.2 %์˜ ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ๊ธฐ๋กํ•˜์˜€๋‹ค. pySTEPS๋Š” ์ƒ๊ด€๊ณ„์ˆ˜์—์„œ 10.2 %, RMSE์™€ NSE์—์„œ ์ตœ๋Œ€ 10.4 %, CSI 5.0์—์„œ ConvLSTM๋ณด๋‹ค ๋†’์€ 10.3 %์˜ ๊ฐœ์„  ํšจ๊ณผ๋ฅผ ๋‚˜ํƒ€๋ƒˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋Š” ์ž…๋ ฅ ๋„๋ฉ”์ธ ํฌ๊ธฐ์˜ ์ฆ๊ฐ€๊ฐ€ ๋‘ ๋ชจ๋ธ์˜ ์˜ˆ์ธก ์„ฑ๋Šฅ์— ๊ธ์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค๋Š” ์ ์„ ๋ณด์—ฌ์ค€๋‹ค.

Fig. 4. Performance of ConvLSTM and pySTEPS with Different Input Domain Size along the Lead Times (R, RMSE, NSE, CSI0.1, CSI1.0, and CSI5.0)

../../Resources/KSCE/Ksce.2025.45.3.0339/fig4.png

Table 1. Improvement of Performance (%) of ConvLSTM and pySTEPS with 256ร—256 and 384ร—384 Input Domains Compared to that with 128ร—128 Input Domain

Model

R

RMSE

NSE

CSI 0.1

CSI 1.0

CSI 5.0

ConvLSTM (256ร—256)

5.2

5.3

5.1

5.5

5.4

5.0

ConvLSTM (384ร—384)

10.4

10.1

10.3

15.2

10.7

10.0

pySTEPS (256ร—256)

5.1

5.0

5.2

5.3

5.2

5.1

pySTEPS (384ร—384)

10.2

10.0

10.4

10.8

10.6

10.3

4. ๊ฒฐ ๋ก 

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋™์ผํ•œ ์•ˆ๋™๋Œ ์œ ์—ญ์— ๋Œ€ํ•ด ์ž…๋ ฅ ๋„๋ฉ”์ธ ํฌ๊ธฐ์˜ ๋ณ€ํ™”์— ๋”ฐ๋ฅธ ConvLSTM๊ณผ pySTEPS ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋ถ„์„ํ•˜์˜€๋‹ค. ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ ๋„๋ฉ”์ธ ํฌ๊ธฐ๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ์˜ˆ์ธก ์ •ํ™•๋„๊ฐ€ ๊ฐœ์„ ๋˜๋Š” ๊ฒฝํ–ฅ์„ ๋ณด์˜€๋‹ค. ํ•˜์ง€๋งŒ ๋ฆฌ๋“œํƒ€์ž„์ด ๊ธธ์–ด์ง์— ๋”ฐ๋ผ ConvLSTM์˜ ์˜ˆ์ธก ์ •ํ™•๋„๋Š” pySTEPS๋ณด๋‹ค ๋น ๋ฅด๊ฒŒ ๊ฐ์†Œํ•˜๋Š” ๊ฒฝํ–ฅ์„ ๋ณด์˜€๋‹ค. ์ด๋Š” ์žฅ๊ธฐ ์‹œ๊ฐ„ ํ”„๋ ˆ์ž„์—์„œ ์ œํ•œ๋œ ๊ณต๊ฐ„ ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๊ฑฐ๋‚˜ ๊ณผ์ ํ•ฉ(overfitting) ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ๊ฐ€๋Šฅ์„ฑ์„ ์‹œ์‚ฌํ•œ๋‹ค. ๋ฐ˜๋ฉด, pySTEPS๋Š” ํ†ต๊ณ„ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๋กœ, 80๋ถ„ ์ด์ƒ์˜ ์žฅ๊ธฐ ๋ฆฌ๋“œ ํƒ€์ž„์—์„œ ๋” ์•ˆ์ •์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค. pySTEPS ๋ชจ๋ธ์€ ConvLSTM๋ณด๋‹ค ์ „๋ฐ˜์ ์œผ๋กœ ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์˜€์œผ๋ฉฐ, ์ด๋Š” ๋‹จ๊ธฐ ๊ฐ•์šฐ ์˜ˆ์ธก์—์„œ pySTEPS๊ฐ€ ๋”์šฑ ์•ˆ์ •์ ์ž„์„ ์‹œ์‚ฌํ•œ๋‹ค.

๋‹จ๊ธฐ ๊ฐ•์šฐ ์˜ˆ์ธก์€ ์‹ค์‹œ๊ฐ„ ํ˜ธ์šฐ๊ฒฝ๋ณด ๋ฐ ์žฌ๋‚œ ๋Œ€์‘ ์ตœ์ ํ™”์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋ฉฐ, ์˜ˆ๋ณด๊ฐ€ ๋น ๋ฅด๊ณ  ์ •ํ™•ํ•  ์ˆ˜๋ก ํšจ๊ณผ์ ์ธ ๋Œ€์‘์ด ๊ฐ€๋Šฅํ•˜๋‹ค. ์ด์— ๋”ฐ๋ผ ๊ฐ€์žฅ ํšจ์œจ์ ์ธ ์˜ˆ๋ณด ๊ฒฝ๋ณด์˜ ์„ ํ–‰ ์‹œ๊ฐ„์— ๋Œ€ํ•œ ๋ถ„์„๊ณผ, ์ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ํšจ์œจ์ ์ธ ์ž…๋ ฅ ๋„๋ฉ”์ธ ์„ค์ • ๋ฐ ๊ฐ•์šฐ ์˜ˆ์ธก ๋ชจ๋ธ ๊ฐœ๋ฐœ์ด ์ค‘์š”ํ•˜๋‹ค. ํ–ฅํ›„ ์—ฐ๊ตฌ๋Š” ๋‹จ๊ธฐ ๊ฐ•์šฐ ์˜ˆ์ธก์˜ ์ •ํ™•๋„๋ฅผ ๋†’์ด๋Š” ๊ฒƒ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ, ์‹ค์‹œ๊ฐ„ ์˜ˆ๋ณด ์‹œ์Šคํ…œ์—์„œ ๋น ๋ฅด๊ณ  ์ •ํ™•ํ•œ ์˜ˆ์ธก์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ๋„๋ก ์ž…๋ ฅ ๋„๋ฉ”์ธ ์ตœ์ ํ™”์™€ ๋ชจ๋ธ ๊ฐœ์„ ์„ ์œ„ํ•œ ์ถ”๊ฐ€์ ์ธ ์—ฐ๊ตฌ๊ฐ€ ์ด๋ฃจ์–ด์ ธ์•ผ ํ•  ๊ฒƒ์ด๋‹ค.

Acknowledgements

This work was supported by the Korea Environment Industry & Technology Institute (KEITI) through the R&D Program for Innovative Flood Protection Technologies against Climate Crisis, funded by the Ministry of Environment (MOE) (No. RS-2023-00218873), and by the Basic Science Research Program through the National Research Foundation of Korea, funded by the Ministry of Science, ICT, and Future Planning (No. RS-2024-00456724).

References

1 
"Carpenter, T. M. and Georgakakos, K. P. (2004). โ€œImpacts of parametric and radar rainfall uncertainty on the ensemble streamflow simulations of a distributed hydrologic model.โ€ Journal of Hydrology, Vol. 298, No. 1-4, pp. 202-221, https://doi.org/10.1016/j.jhydrol.2004.03.036."DOI
2 
"Choi, S. and Kim, Y. (2022). โ€œRad-cGAN v1.0: Radar-based precipitation nowcasting model with conditional generative adversarial networks for multiple dam domains.โ€ Geoscientific Model Development, Vol. 15, No. 15, pp. 5967-5985, https://doi.org/10.5194/gmd-15-5967-2022."DOI
3 
"Foresti, L., Reyniers, M., Seed, A. and Delobbe, L. (2016). โ€œDevelopment and verification of a real-time stochastic precipitation nowcasting system for urban hydrology in Belgium.โ€ Hydrology and Earth System Sciences, Vol. 20, No. 1, pp. 505-527, https://doi.org/10.5194/hess-20-505-2016."DOI
4 
"Giannone, D., Reichlin, L. and Small, D. (2008). โ€œNowcasting: The real-time informational content of macroeconomic data.โ€ Journal of Monetary Economics, Vol. 55, No. 4, pp. 665-676, https://doi.org/10.1016/j.jmoneco.2008.05.010."DOI
5 
"Heuvelink, D., Berenguer, M., Brauer, C. C. and Uijlenhoet, R. (2020). โ€œHydrological application of radar rainfall nowcasting in the Netherlands.โ€ Environment International, Vol. 136, p. 105431, https://doi.org/10.1016/j.envint.2019.105431."DOI
6 
"Kim, S., Hong, S., Joh, M. and Song, S. K. (2017). โ€œDeepRain: ConvLSTM network for precipitation prediction using multichannel radar data.โ€ arXiv preprint, arXiv:1711.02316, https://arxiv.org/abs/1711.02316."DOI
7 
"Marshall, J. S. and Palmer, W. M. K. (1948). โ€œThe distribution of raindrops with size.โ€ Journal of Atmospheric Sciences, Vol. 5, No. 4, pp. 165-166, https://doi.org/10.1175/1520-0469(1948)005<0165:TDORWS>2.0.CO;2."DOI
8 
"Pulkkinen, S., Nerini, D., Pรฉrez Hortal, A. A., Velasco-Forero, C., Seed, A., Germann, U. and Foresti, L. (2019). โ€œPysteps: An open-source Python library for probabilistic precipitation nowcasting (v1.0).โ€ Geoscientific Model Development, Vol. 12, No. 12, pp. 4185-4219, https://doi.org/10.5194/gmd-12-4185-2019."DOI
9 
"Shi, X., Chen, Z., Wang, H., Yeung, D. Y., Wong, W. K. and Woo, W. C. (2015). โ€œConvolutional LSTM network: A machine learning approach for precipitation nowcasting.โ€ Advances in Neural Information Processing Systems, Vol. 28, https://arxiv.org/abs/1506.04214."DOI
10 
"Smith, J., Birch, C., Marsham, J., Peatman, S., Bollasina, M. and Pankiewicz, G. (2024). โ€œEvaluating pySTEPS optical flow algorithms for convection nowcasting over the Maritime Continent using satellite data.โ€ Natural Hazards and Earth System Sciences, Vol. 24, No. 2, pp. 567-582, https://doi.org/10.5194/nhess-24-567-2024."DOI