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

  1. ํ•œ๊ตญ๊ตํ†ต์—ฐ๊ตฌ์› ์—ฐ๊ตฌ์›, ๊ณตํ•™์„์‚ฌ (The Korea Transport Institute)
  2. ํ•œ๋ฐญ๋Œ€ํ•™๊ต ๋„์‹œ๊ณตํ•™๊ณผ ๊ต์ˆ˜, ๊ณตํ•™๋ฐ•์‚ฌ (Hanbat National University)


๋ฌด์ธํ•ญ๊ณต๊ธฐ, ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘, ์ด๋ฏธ์ง€ ๋ถ„ํ• ๋ฐฉ๋ฒ•, ์ด๋ฏธ์ง€ ๊ฒ€์ถœ๋ฐฉ๋ฒ•, ๋”ฅ๋Ÿฌ๋‹
UAV, Pine wilt disease, YOLOv2, SegNet, Deep learning

  • 1. ์„œ ๋ก 

  • 2. ๊ธฐ์กด์—ฐ๊ตฌ๊ณ ์ฐฐ

  • 3. ํ•ญ๊ณต์˜์ƒ์ดฌ์˜์„ ํ†ตํ•œ ์ •์‚ฌ์˜์ƒ ์ƒ์„ฑ

  •   3.1 ์—ฐ๊ตฌ ๋Œ€์ƒ์ง€ ์„ ์ •

  •   3.2 ์˜์ƒ๋ฐ์ดํ„ฐ ํš๋“ ๋ฐ ์ •์‚ฌ์˜์ƒ ์ œ์ž‘

  • 4. ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜์˜ ๊ฐ์ฒด ์ธ์‹๊ณผ ํŒ๋ณ„

  •   4.1 SegNet๊ณผ YOLO ๊ธฐ๋ฐ˜ ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜

  •   4.2 ํ•™์Šต ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ๊ณผ ํŒŒ๋ผ๋ฉ”ํƒ€ ์„ค์ •

  •   4.3 ๋ถ„์„ ๊ฒฐ๊ณผ

  • 5. ๊ฒฐ ๋ก 

1. ์„œ ๋ก 

์šฐ๋ฆฌ๋‚˜๋ผ์˜ ์†Œ๋‚˜๋ฌด๋Š” ๊ณผ๊ฑฐ๋ถ€ํ„ฐ ์“ธ๋ชจ๊ฐ€ ๋งŽ๊ณ  ์„ฑ์žฅํ™˜๊ฒฝ์ด ์ ํ•ฉํ•˜์—ฌ ์ •์ฑ…์ ์œผ๋กœ ๋ณดํ˜ธํ•˜๊ณ  ๋ฌด๋ถ„๋ณ„ํ•œ ๋ฒŒ์ฑ„๋ฅผ ๊ธˆ์ง€ํ•˜๋Š” ๋“ฑ์˜ ๊ด€๋ฆฌ๋ฅผ ํ•˜๊ณ  ์žˆ๋‹ค. ํ•˜์ง€๋งŒ, ์ด๋Ÿฐ ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ตœ๊ทผ๊นŒ์ง€ ์‚ฐ๋ฆผ์˜ ๋งŽ์€ ๋ถ€๋ถ„์„ ์ฐจ์ง€ํ•˜๋˜ ์†Œ๋‚˜๋ฌด๊ฐ€ ๊ธ‰๊ฒฉํ•˜๊ฒŒ ์ค„์–ด๋“ค๊ณ  ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜„์ƒ์˜ ๊ฐ€์žฅ ํฐ ์›์ธ์€ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘์˜ ์ „์—ผ ๋•Œ๋ฌธ์ด๋ฉฐ, ํ•œ๋ฒˆ ๊ฐ์—ผ๋˜๋ฉด ๋ฟŒ๋ฆฌ๋กœ๋ถ€ํ„ฐ ๊ณต๊ธ‰๋˜๋Š” ์ˆ˜๋ถ„๊ณผ ์–‘๋ถ„์˜ ์ด๋™์„ ๋ฐฉํ•ดํ•˜์—ฌ ์žŽ์ด ์ฒ˜์ง€๊ณ  ๊ฐˆ๋ณ€ํ•˜์—ฌ ๊ฐ์—ผ ํ›„ 20์ผ ์ •๋„๊ฐ€ ์ง€๋‚˜๊ฒŒ ๋˜๋ฉด ๊ฒฐ๊ตญ์—๋Š” 100 % ๊ณ ์‚ฌํ•˜๋Š” ์‹ฌ๊ฐํ•œ ๋ณ‘์ด๋‹ค.

์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘์€ ์ดˆ๊ธฐ ๋ฐฉ์ œ์˜ ๋ฏธํก์œผ๋กœ ๊ฒฝ์ƒ๋‚จยท๋ถ๋„๋ฅผ ์‹œ์ž‘ํ•ด ์ „๊ตญ์ ์œผ๋กœ ํ”ผํ•ด๊ฐ€ ๊ธ‰์ฆํ•˜๊ฒŒ ๋˜์—ˆ๊ณ  ์‚ฌํƒœ์˜ ์‹ฌ๊ฐ์„ฑ์„ ๋Š๋‚€ ์ •๋ถ€์—์„œ 2005๋…„ ๏ฝข์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๋ฐฉ์ œํŠน๋ณ„๋ฒ•๏ฝฃ์„ ์‹œํ–‰ํ•˜์—ฌ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘์œผ๋กœ๋ถ€ํ„ฐ ํ”ผํ•ด๋ฅผ ๋ฐ›๋Š” ์‚ฐ๋ฆผ์„ ๋ณดํ˜ธํ•˜๊ณ , ์‚ฐ๋ฆผ์ž์›์œผ๋กœ์จ์˜ ๊ธฐ๋Šฅ์„ ํ™•๋ณดํ•˜๊ธฐ ์œ„ํ•œ ํ”ผํ•ด ๋ฐฉ์ง€ ๋Œ€์ฑ…์„ ๊ฐ•๊ตฌํ•˜๊ณ  ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๊ฐ€๋ญ„, ๊ณ ์˜จํ˜„์ƒ ๋“ฑ์˜ ๊ธฐํ›„์  ์š”์ธ๊ณผ ๊ณ ์‚ฌ๋ชฉ์˜ ๋ฐฉ์น˜, ๊ฐ์—ผ๋ชฉ์˜ ๋ฌด๋‹จ์ด๋™ ๋“ฑ์œผ๋กœ ํ”ผํ•ด ๊ทœ๋ชจ๋Š” ์ ์ฐจ ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ๋‹ค. ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด์™€ ํ™•์‚ฐ์„ ๋ง‰๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ ์ธ๋ ฅ์— ์˜ํ•œ ํ˜„์ง€์กฐ์‚ฌ๋ฅผ ์‹ค์‹œํ•˜๊ณ  ์กฐ๊ธฐ๋ฐœ๊ฒฌ์„ ์œ„ํ•ด ์‚ฐ๋ฆผ์ง€์—ญ์— ์˜ˆ์ฐฐ์›์„ ๋ฐฐ์น˜ํ•˜์—ฌ ๋ฐœ์ƒ์ง€์—ญ์˜ ๊ฐ์—ผ๋ชฉ์„ ์‹ ์†ํ•˜๊ฒŒ ์ œ๊ฑฐํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋ฉฐ ๋ฐ˜์ถœ๊ธˆ์ง€๊ตฌ์—ญ์„ ์„ค์ •ํ•˜๋Š” ๋“ฑ์˜ ๊ด€๋ฆฌ์— ํž˜์“ฐ๊ณ  ์žˆ์œผ๋‚˜, ํ›ˆ์ฆ ์ฒ˜๋ฆฌ๋กœ ์ด๋ฃจ์–ด์ง€๋Š” ์‚ฌํ›„๋Œ€์‘ํ˜• ๋ฐฉ์‹์˜ ํ•œ๊ณ„๋ฅผ ๋ณด์ด๊ณ  ์žˆ๋‹ค(Kim et al., 2016).

ํŠนํžˆ, ์ธ๋ ฅ ๊ธฐ๋ฐ˜์˜ ์ง€์ƒ์˜ˆ์ฐฐ ๋ฐฉ๋ฒ•์€ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘์˜ ๋™์‹œ๋‹ค๋ฐœ์ ์ด๊ณ  ๊ธ‰์ง„์ ์ธ ๋ฐœ์ƒ ํŒจํ„ด๊ณผ ์‚ฌ๋žŒ์ด ์ ‘๊ทผํ•˜๊ธฐ ์–ด๋ ค์šด ๊ณ ์‚ฐ์ง€, ๊ธ‰๊ฒฝ์‚ฌ์ง€, ์ ˆ๋ฒฝ ๊ฐ™์€ ๊ณณ์—๋„ ํ”ผํ•ด ๋ฐœ์ƒ ์ดˆ๊ธฐ์—๋Š” ์ ์ƒ์œผ๋กœ ๋ฐœ์ƒํ•˜์—ฌ ๊ด€์ธก์ด ์–ด๋ ต๋‹ค๋Š” ์  ๋“ฑ์„ ๊ฐ์•ˆํ•˜๋ฉด ์‹ ์†ํ•œ ํ”ผํ•ด ์ง€์—ญ ์œ„์น˜ ํŒŒ์•…๊ณผ ์ •ํ™•ํ•œ ์˜ˆ์ธก์— ํ•œ๊ณ„๊ฐ€ ์žˆ๋‹ค.

์ตœ๊ทผ 4์ฐจ ์‚ฐ์—…ํ˜๋ช…์˜ ํ•ต์‹ฌ๊ธฐ์ˆ  ๊ฐ€์šด๋ฐ ๋ฌด์ธํ•ญ๊ณต๊ธฐ(Unmanned Aerial Vehicle, UAV)์˜ ํ•ญ๊ณต์ดฌ์˜ ๊ธฐ์ˆ ์€ ์‚ฌ๋žŒ์ด ์ ‘๊ทผํ•˜๊ธฐ ํž˜๋“  ์ง€์—ญ์„ ์‹ค์‹œ๊ฐ„์œผ๋กœ ๋ชจ๋‹ˆํ„ฐ๋งํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ์ €๋ ดํ•œ ๋น„์šฉ์œผ๋กœ ๋ฐ์ดํ„ฐ ํ™•๋ณด๊ฐ€ ๊ฐ€๋Šฅํ•ด ๋‹ค์–‘ํ•œ ๋ถ„์•ผ์—์„œ ํšจ์œจ์ ์œผ๋กœ ํ™œ์šฉ๋˜๊ณ  ์žˆ๋‹ค. ํŠนํžˆ ์ด์™€ ๊ด€๋ จํ•œ ์œตํ•ฉ ๊ธฐ์ˆ ๋กœ ์˜์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•œ ๊ณต๊ฐ„์ •๋ณด๋ถ„์„ ๊ธฐ์ˆ ์€ ์‚ฐ๋ฆผ, ๋ฐฉ์žฌ, ํ™˜๊ฒฝ ๋ถ„์•ผ์—์„œ๋„ ์ ๊ทน์ ์œผ๋กœ ๋„์ž…๋˜๊ณ  ์žˆ๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ธฐ์กด์˜ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ ํƒ์ง€ ๋ฐฉ๋ฒ•์˜ ํ•œ๊ณ„๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋ฌด์ธํ•ญ๊ณต๊ธฐ๋ฅผ ์ด์šฉํ•œ ์˜์ƒ ์ž๋ฃŒ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋”ฅ๋Ÿฌ๋‹(Deep Learning) ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ด์šฉํ•˜์—ฌ ์ž๋™ ์ธ์‹ํ•˜๋Š” ๋ฐฉ๋ฒ•๋ก ์„ ์ œ์‹œํ•˜๊ณ  ์‹ค์ œ ํ˜„์žฅ์—์„œ์˜ ํ™œ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ํ™•์ธํ•˜๋Š” ๊ฒƒ์„ ๋ชฉ์ ์œผ๋กœ ํ•œ๋‹ค.

2. ๊ธฐ์กด์—ฐ๊ตฌ๊ณ ์ฐฐ

์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ์„ ํ™•์ธํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ๊ธฐ์กด์˜ ์ธ๋ ฅ ๊ธฐ๋ฐ˜ ํ˜„์žฅ์กฐ์‚ฌ ์ด์™ธ์— ํ—ฌ๊ธฐ๋ฅผ ํƒ€๊ณ  ํ•ญ๊ณต์˜ˆ์ฐฐ์„ ์‹ค์‹œํ•˜๊ฑฐ๋‚˜, ์œ„์„ฑ์˜์ƒ์„ ์ด์šฉํ•œ ํ”ผํ•ด ์ƒํ™ฉ ํŒŒ์•…๊ณผ ๊ฐ™์€ ์—ฐ๊ตฌ๊ฐ€ ์ง„ํ–‰๋˜์—ˆ๋‹ค. Son et al.(2006)์€ ์œ„์„ฑ์˜์ƒ์„ ํ™œ์šฉํ•˜์—ฌ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด ์ง€์—ญ์„ ์ถ”์ถœํ•˜๋Š” ๊ธฐ์ดˆ์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•˜์˜€๊ณ , Kim et al.(2001)์€ ์œ„์„ฑ์˜์ƒ์„ ๋ฐ”ํƒ•์œผ๋กœ GIS ํ”„๋กœ๊ทธ๋žจ์—์„œ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด ์ง€์—ญ์„ ์ถ”์ถœํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•œ ๋ฐ” ์žˆ๋‹ค.

๊ทธ๋Ÿฌ๋‚˜ ์œ„์„ฑ์˜์ƒ์˜ ๊ฒฝ์šฐ๋Š” ๊ณ ํ•ด์ƒ๋„์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ™•๋ณดํ•˜๊ธฐ ์œ„ํ•ด ์ˆ˜๋ฐ˜๋˜๋Š” ๋น„์šฉ ๋ฌธ์ œ ์ด์™ธ์—๋„ ๊ธฐ์ƒ ์กฐ๊ฑด์— ๋”ฐ๋ฅธ ๋ถ„์„ ๊ฐ€๋Šฅ ์—ฌ๋ถ€์™€ ์ฆ‰๊ฐ์ ์ธ ๋Œ€์ฒ˜๋ฅผ ์œ„ํ•œ ๋ถ„์„์— ํ•œ๊ณ„๊ฐ€ ์žˆ์—ˆ๋‹ค. ์ด๋ณด๋‹ค ๋น ๋ฅธ ๋Œ€์‘์„ ์œ„ํ•ด์„œ Kim et al.(2010)์€ ํ—ฌ๊ธฐ๋ฅผ ํƒ€๊ณ ์„œ ํ•ญ๊ณต์ •๋ฐ€์˜ˆ์ฐฐ์„ ์‹ค์‹œํ•˜์˜€์œผ๋ฉฐ, GIS ํ”„๋กœ๊ทธ๋žจ์„ ํ™œ์šฉํ•˜์—ฌ ์ •๋ฐ€ ๋ถ„์„์„ ์ง„ํ–‰ํ•˜์˜€์œผ๋‚˜, ํ—ฌ๊ธฐ์˜ 1ํšŒ ๋น„ํ–‰์— ์†Œ๋น„๋˜๋Š” ๋น„์šฉ๊ณผ ์กฐ์‚ฌ ๋ฐฉ๋ฒ•์˜ ํŠน์„ฑ์ƒ ์œก์•ˆ์œผ๋กœ ๊ด€์ธกํ•œ ์ •๋ณด๋ฅผ ์ง์ ‘ ์•ผ์žฅ์— ๊ธฐ์ž…ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํ”ผํ•ด๋ชฉ ์œ„์น˜๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ •ํ™•ํ•œ ์ขŒํ‘œ๋ฅผ ์–ป๊ธฐ์—๋Š” ์–ด๋ ค์›€์ด ์žˆ์–ด์„œ ํ˜„ํ™ฉ ํŒŒ์•… ์ˆ˜์ค€์˜ ์ž๋ฃŒ์ˆ˜์ง‘์ด๋ผ๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์—ˆ๋‹ค.

4์ฐจ ์‚ฐ์—…ํ˜๋ช…์˜ ํ•ต์‹ฌ๊ธฐ์ˆ ์ธ ๋ฌด์ธํ•ญ๊ณต๊ธฐ์˜ ํ•ญ๊ณต์ดฌ์˜ ๊ธฐ์ˆ ์€ ๋‹ค์–‘ํ•œ ๊ณณ์„ ์‹ค์‹œ๊ฐ„์œผ๋กœ ๋ชจ๋‹ˆํ„ฐ๋ง ๊ฐ€๋Šฅํ•˜๋ฉฐ ์†์‰ฝ๊ฒŒ ๋ฐ์ดํ„ฐ ํ™•๋ณด๊ฐ€ ๊ฐ€๋Šฅํ•ด ๋งค์šฐ ํšจ์œจ์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ๋‹ค. ํŠนํžˆ, ์ด์™€ ๊ด€๋ จํ•œ ์œตํ•ฉ ๊ธฐ์ˆ ๋กœ ์˜์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•œ ๊ณต๊ฐ„์ •๋ณด๋ถ„์„ ๊ธฐ์ˆ ์€ ์‚ฐ๋ฆผ ๋ถ„์•ผ์—๋„ ๋„๋ฆฌ ๋„์ž…๋˜๊ณ  ์žˆ๋‹ค(Han, 2019).

์‚ฐ๋ฆผํ”ผํ•ด๋ฅผ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•ด Kim and Kim(2008)์€ ๋ฌด์ธํ•ญ๊ณต๊ธฐ๋ฅผ ์ด์šฉํ•œ ์ดฌ์˜๋ฐฉ๋ฒ•, ์ •์‚ฌ์˜์ƒ(orthomosaic) ์ œ์ž‘, ์˜์ƒ๋ถ„๋ฅ˜ ๋ฐ ํ•„ํ„ฐ๋ง, ํ˜„์žฅ์กฐ์‚ฌ๋ฅผ ํ†ตํ•œ ์˜ˆ์ฐฐ ๋ฐฉ๋ฒ•๋ก ์„ ์ œ์‹œํ•˜์˜€๋‹ค. ํŠนํžˆ, ๋ถ„์„์„ ์œ„ํ•œ ๊ฐ์ฒด ์œ„์น˜๋‚˜ ์ขŒํ‘œ๋ฅผ ํš๋“ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ํ•ญ๊ณต์ดฌ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•œ ํ•ญ๊ณต์‚ผ๊ฐ์ธก๋Ÿ‰(aerotriangulation)๊ณผ ์ •์‚ฌ์˜์ƒ ์ œ์ž‘ ๊ณผ์ •์ด ํ•„์ˆ˜๋กœ ์ง„ํ–‰๋˜์–ด์•ผ ํ•จ์„ ๊ฐ•์กฐํ•˜์˜€๋‹ค.

์ •๋„ ๋†’์€ ๋ถ„์„์„ ์œ„ํ•ด์„œ๋Š” ํ•ญ๊ณต์ดฌ์˜ ์‹œ ๊ธฐ๋ก๋˜๋Š” ์œ„์น˜์ •๋ณด(Global Positioning System, GPS) ์™ธ์— ์ง€์ƒํ‘œ๋ณธ๊ฑฐ๋ฆฌ(Ground Sampling Distance, GSD)๋ฅผ ์ธก์ •ํ•˜๊ณ  ๋ถ„์„์— ํฌํ•จํ•  ๊ฒฝ์šฐ ๊ธฐ์กด์˜ ๊ฒฐ๊ณผ ๋ณด๋‹ค ๊ฐ’์˜ ์ •๋„๊ฐ€ ํ–ฅ์ƒ๋˜๋ฉฐ, ์–ป๊ณ ์ž ํ•˜๋Š” ์œ„์น˜์ •๋ณด์˜ ์‹ค์ œ ๊ฐ’์— ๊ฐ€๊นŒ์šด ์ •๋ณด๋ฅผ ํš๋“ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ฐํžˆ๊ธฐ๋„ ํ•˜์˜€๋‹ค(Do et al., 2018; Kim et al., 2014).

Nagai et al.(2009)์€ ๋ฌด์ธํ—ฌ๊ธฐ์— ์นด๋ฉ”๋ผ๋ฅผ ์žฅ์ฐฉํ•˜์—ฌ ํ•ด์•ˆ ์‚ฐ๋ฆผ์— ๋Œ€ํ•œ ์ง€๋ฆฌ์  ์œ„์น˜๋ฅผ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ๋กœ 10~30 cm ์ •๋„์˜ ์˜ค์ฐจ ๋ฒ”์œ„์—์„œ ๊ฐ€๋Šฅํ•จ์„ ๋ฐํ˜”๋‹ค. Rokhmana(2015)๋Š” ๋ฌด์ธํ•ญ๊ณต๊ธฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‚ฐ๋ฆผ๊ณผ ๋†์ง€๋ฅผ ์ดฌ์˜ํ•œ ์˜์ƒ ๋ฐ์ดํ„ฐ์˜ ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ๋กœ ์ˆ˜ํ‰์œผ๋กœ 2ํ”ฝ์…€, ์ˆ˜์ง์œผ๋กœ 5ํ”ฝ์…€์˜ ์˜ค์ฐจ๊ฐ€ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ 3D ๊ณต๊ฐ„๋ถ„์„์œผ๋กœ ์ˆ˜๋ชฉ์˜ ๋†’์ด๋„ ์ธก์ •ํ•  ์ˆ˜ ์žˆ์Œ์„ ํ™•์ธํ•˜์˜€๋‹ค.

Kim et al.(2017)์€ ๋ฌด์ธํ•ญ๊ณต๊ธฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์„ธ์ข…์‹œ ์ฃผ๋ณ€์˜ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด์ง€๋ฅผ ๋ถ„์„ํ•˜๊ณ  ํ”ผํ•ด๋ชฉ์˜ ๋ถ„ํฌ ํŠน์„ฑ์„ ๋ถ„์„ํ•˜์˜€์œผ๋ฉฐ, Lee(2017)๋Š” ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด๋ชฉ์„ ๋ฌด์ธํ•ญ๊ณต๊ธฐ ์˜์ƒ์œผ๋กœ ์ดฌ์˜ํ•˜์—ฌ GIS ํ”„๋กœ๊ทธ๋žจ์—์„œ RGB ๊ฐ’ ๊ธฐ์ค€์œผ๋กœ ํ”ผํ•ด๋ชฉ์˜ ์œ„์น˜๋ฅผ ํ™”์†Œ ๋‹จ์œ„๋กœ ์ถ”์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•๋ก ์„ ์ œ์‹œํ•˜์˜€๋‹ค.

ํ•œํŽธ ๋ฌด์ธํ•ญ๊ณต๊ธฐ์— ๋‹ค์–‘ํ•œ ์นด๋ฉ”๋ผ ํ˜น์€ ์„ผ์„œ๋ฅผ ์žฅ์ฐฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•˜๋Š” ์—ฐ๊ตฌ๊ฐ€ ๋งŽ์ด ์ง„ํ–‰๋˜์—ˆ๋‹ค. Lee et al.(2014)์€ ์ดˆ๋ถ„๊ด‘ ์˜์ƒ ์ดฌ์˜์„ ํ™œ์šฉํ•˜์—ฌ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘์— ์˜ํ•œ ๊ฐ์—ผ๋ชฉ ํƒ์ง€์™€ ๊ฐ์—ผ๋ชฉ์ด ๋‚˜ํƒ€๋‚ด๋Š” ๋ถ„๊ด‘ ํŠน์„ฑ์„ ๋ถ„์„ํ•˜์—ฌ ์‹œ๋“ฆ ํ˜„์ƒ์ด ์œก์•ˆ์œผ๋กœ ๊ด€์ธก๋˜๋Š” ์‹œ์ ์ด ๊ฐ์—ผ ํ›„ 2๊ฐœ์›” ํ›„์— ๊ด€์ฐฐ ๊ฐ€๋Šฅํ•จ์„ ๋ณด์˜€๋‹ค. ๋˜ํ•œ, Kim et al.(2015)์€ ์ดˆ๋ถ„๊ด‘ ์˜์ƒ ์ดฌ์˜์„ ์‹œ๊ธฐ๋ณ„๋กœ ์ง„ํ–‰ํ•˜๊ณ  ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด๋ชฉ์˜ ์‹์ƒ์ง€์ˆ˜๋ฅผ ์‹œ๊ณ„์—ด ๋ณ„๋กœ ์ถ”์ ํ•˜์—ฌ ํ”ผํ•ด๋ชฉ์˜ ์œ„์น˜์™€ ๋ฒ”์œ„๋ฅผ ๋ถ„์„ํ•˜์˜€์œผ๋ฉฐ, ์†Œ๋‚˜๋ฌด๋ฅผ ๋Œ€์ƒ์œผ๋กœ ์ œ์ดˆ์ œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ๋“ฆ ํ˜„์ƒ์— ๋”ฐ๋ผ์„œ ์‹์ƒ์ง€์ˆ˜์˜ ๋ชจ๋ธ์ด ์–ด๋–ป๊ฒŒ ๋‚˜ํƒ€๋‚˜๋Š”์ง€ ๋ถ„์„ํ•˜์˜€๋‹ค(Dash et al., 2017).

Xu et al.(2018)์€ ๋‹ค์ค‘๋ถ„๊ด‘ ์นด๋ฉ”๋ผ๋ฅผ ๋ฌด์ธํ•ญ๊ณต๊ธฐ์— ํƒ‘์žฌํ•˜์—ฌ ํ† ์ง€ํ”ผ๋ณต๋„๋ฅผ ์ œ์ž‘ํ•œ ๊ฒฐ๊ณผ ์˜ค์ฐจ๊ฐ€ ๋งค์šฐ ์ž‘์€ RGB ์˜์ƒ๊ณผ ๋‹ค์ค‘๋ถ„๊ด‘์˜์ƒ์„ ์ œ์ž‘ํ•  ์ˆ˜ ์žˆ์—ˆ์œผ๋ฉฐ, ์ดฌ์˜ ์‹œ๊ธฐ๋ณ„๋กœ ์ •ํ™•๋„๋ฅผ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ ๋งค์šฐ ๋†’์€ ์ •ํ™•๋„๊ฐ€ ๋‚˜ํƒ€๋‚˜ ๋‹ค์ค‘๋ถ„๊ด‘ ์นด๋ฉ”๋ผ์˜ ํ™œ์šฉ์„ฑ์„ ํ™•์ธํ•˜์˜€๋‹ค.

Smigaj et al.(2019)์€ ๋ณ‘์— ๊ฑธ๋ฆฐ ์†Œ๋‚˜๋ฌด์— ์—ดํ™”์ƒ ์นด๋ฉ”๋ผ์™€ ๋ผ์ด๋‹ค(Light Detection and Ranging, LiDAR) ์„ผ์„œ๋ฅผ ์ด์šฉํ•˜์—ฌ ์ˆ˜๋ชฉ์˜ ์˜จ๋„ ๋ณ€ํ™”๋ฅผ ํ™•์ธํ•˜์˜€๊ณ , ๋‚˜๋ฌด๊ฐ€ ์•„๋‹Œ ๋ฐ€์ด๋‚˜ ๋ฌด์™€ ๊ฐ™์€ ๋†์ž‘๋ฌผ์—๋„ ์‹์ƒ์ง€์ˆ˜ ๋ณ€ํ™”๊ฐ€ ์žˆ์Œ์„ ๋ฐํ˜”๋‹ค(Dash et al., 2017; Im et al., 2018).

์ตœ๊ทผ์—๋Š” ์ธ๊ณต์ง€๋Šฅ์˜ ๋ฐœ๋‹ฌ๊ณผ ๋‹ค์–‘ํ•œ ์ •๋ณด์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ๊ฐ€ ์ถ•์ ๋˜๋ฉด์„œ ์ด๋ฅผ ์ž๋™์œผ๋กœ ๋น ๋ฅด๊ฒŒ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ์ˆ ๋“ค์ด ๊ฐœ๋ฐœ๋˜๊ณ  ์žˆ๋‹ค. Badrinarayanan et al.(2017)์€ ๊ธฐ์กด ์˜์ƒ ๋ถ„์„ ๋ฐฉ์‹์— ๋ฒ—์–ด๋‚˜ ์ƒˆ๋กญ๊ณ  ์‹ค์šฉ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ํ”ฝ์…€ ํ•˜๋‚˜์— ์˜๋ฏธ๋ฅผ ๋ถ€์—ฌํ•˜๋Š” ์ด๋ฏธ์ง€ ๋ถ„ํ• ๋ฐฉ์‹์ธ SegNet ๊ตฌ์กฐ๋ฅผ ์—ฐ๊ตฌํ•˜๊ณ  ์˜์ƒ๋ถ„์„์— ๋Œ€ํ•œ ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ํ•˜์˜€๋‹ค. ์ด์™€๋Š” ์กฐ๊ธˆ ๋‹ค๋ฅธ ์ ‘๊ทผ๋ฐฉ์‹์œผ๋กœ Redmon et al.(2016)์€ ๊ฐ์ฒด ํƒ์ง€์˜ ์ƒˆ๋กœ์šด ๊ตฌ์กฐ๋กœ YOLO๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๊ธฐ์กด์˜ ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•œ ํƒ์ง€ ๋ฐฉ์•ˆ์„ ์ œ์‹œํ•˜๊ธฐ๋„ ํ•˜์˜€๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํƒ์ง€ ์„ฑ๋Šฅ ํ‰๊ฐ€ ์ค‘ ์žฌํ˜„์œจ์ด ๋–จ์–ด์ง€๋Š” ๋ฌธ์ œ๊ฐ€ ๋‚˜ํƒ€๋‚˜์ž ์ดํ›„ Redmon and Farhadi(2017)์€ ์†๋„์™€ ์ •ํ™•๋„๊ฐ€ ๋†’์œผ๋ฉด์„œ ๋™์‹œ์— ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ๊ฐ์ฒด๋ฅผ ์‹ค์‹œ๊ฐ„์œผ๋กœ ํƒ์ง€ํ•  ์ˆ˜ ์žˆ๋Š” YOLOv2๋ฅผ ํ™œ์šฉํ•œ ์—ฐ๊ตฌ๋ฅผ ์ œ์‹œํ•˜์˜€๋‹ค(Giuffrida et al., 2019).

์ผ๋ฐ˜์ ์œผ๋กœ ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ์ˆ ์€ ์ •ํ™•๋„๊ฐ€ ๋†’์œผ๋ฉด ํšจ์œจ์„ฑ์ด ๋›ฐ์–ด๋‚˜๋‹ค๊ณ  ํ•ด์„ํ•˜์ง€๋งŒ, ์‹ค์ œ๋กœ ๋ฐ์ดํ„ฐ ํŠน์„ฑ์— ๋”ฐ๋ผ์„œ ๋ถ„์„๋œ ๋‚ด์šฉ์ด ์ ํ•ฉํ•œ์ง€๋ฅผ ์˜์‚ฌ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๊ฐ์ฒด ์ง‘ํ•ฉ ๊ฐ„์˜ ์ฐจ์ด๋ฅผ ํ™•์ธํ•˜๊ณ , ๋ถˆ๊ท ํ˜•ํ•œ ๋ฐ์ดํ„ฐ์˜ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์„ฑ๋Šฅ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์ด ํ•„์š”ํ•˜๋‹ค(Derczynski, 2016; Kotsiantis et al., 2006).

ํŠนํžˆ, ๊ด‘๋ฒ”์œ„ํ•œ ์ง€์—ญ์„ ๋Œ€์ƒ์œผ๋กœ ํ•˜๋Š” ์‚ฐ๋ฆผ์ •๋ณด๋ฅผ ๋ถ„์„ํ•˜๊ธฐ ์œ„ํ•œ ์—ฐ๊ตฌ๋กœ ์ตœ๊ทผ์—๋Š” ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ ์‚ฐ๋ฆผ ๋ฐ์ดํ„ฐ๋ฅผ ์ž๋™์œผ๋กœ ๋ถ„๋ฅ˜ํ•ด ์ฃผ๋Š” ์—ฐ๊ตฌ๊ฐ€ ํ™œ๋ฐœํ•˜๊ฒŒ ์ง„ํ–‰ ์ค‘์— ์žˆ๋‹ค. Dyson et al.(2019)์€ NDVI์™€ ๊ณ ํ•ด์ƒ๋„ DSM ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šต์‹œ์ผœ ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง(Convolutional Neural Network, CNN) ๊ธฐ๋ฐ˜์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ™œ์šฉํ•˜์—ฌ ํ† ์–‘๊ณผ ๋†์ž‘๋ฌผ์— ๋Œ€ํ•ด ๋ถ„ํ• ์ด ๊ฐ€๋Šฅํ•จ์„ ๋ฐํ˜”๊ณ , Lee et al.(2019)์€ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด๋ชฉ์„ ๋จธ์‹ ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜์˜ ์ธ๊ณต์‹ ๊ฒฝ๋ง์„ ํ™œ์šฉํ•˜์—ฌ ํƒ์ง€ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ฐํ˜”๋‹ค.

3. ํ•ญ๊ณต์˜์ƒ์ดฌ์˜์„ ํ†ตํ•œ ์ •์‚ฌ์˜์ƒ ์ƒ์„ฑ

๋ณธ ์ ˆ์—์„œ๋Š” ๊ธฐ์กด์˜ ํ˜„์žฅ์— ํˆฌ์ž…๋˜๋Š” ์ธ๋ ฅ์„ ํ†ตํ•œ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ ์˜ˆ์ฐฐ ๋ฐฉ๋ฒ•์˜ ๋ฌธ์ œ์ ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋ฌด์ธํ•ญ๊ณต๊ธฐ๋ฅผ ํ™œ์šฉํ•œ ์‚ฐ๋ฆผ์ง€์—ญ์˜ ์˜์ƒ ์ž๋ฃŒ ์ทจ๋“๊ณผ ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜์˜ ๊ฐ์ฒด ์ธ์‹ ํƒ์ง€ ๊ธฐ๋ฒ•์˜ ๋ฐฉ๋ฒ•๋ก ์„ ์ œ์‹œํ•˜๊ธฐ๋กœ ํ•œ๋‹ค.

3.1 ์—ฐ๊ตฌ ๋Œ€์ƒ์ง€ ์„ ์ •

๋จผ์ €, ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ์˜ ๋ถ„์„๊ณผ ๊ฒ€์ฆ์„ ์œ„ํ•ด ๋Œ€์ƒ์ง€๋ฅผ ์„ ์ •ํ•˜์˜€๋‹ค. ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ๋Œ€์ƒ์ง€๋Š” ๊ฒฝ์ƒ๋ถ๋„ ํฌํ•ญ์‹œ ๋ถ๊ตฌ ๊ธฐ๊ณ„๋ฉด ๋ด‰๊ฐ•์žฌ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ๋ถ๋™์ชฝ ์‚ฐ์ง€์™€ ์ง€๋ฆฌ์ƒ ๋‚จ์ชฝ์— ์œ„์น˜ํ•˜๋Š” ๋งˆ๋ด‰์‚ฐ์œผ๋กœ ๋‘ ์‚ฐ์˜ ์ตœ๋Œ€ ๋†’์ด 209.3 m๋กœ ๋น„๊ต์  ๋‚ฎ์€ ์‚ฐ์„ธ์ด๋ฉฐ ๋“ฑ์‚ฐ ์ฝ”์Šค์™€ ์‚ฌ๋žŒ์ด ๋‹ค๋‹ ์ˆ˜ ์žˆ๋Š” ์ž„๋„๊ฐ€ ์žˆ์–ด ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ ์—ฌ๋ถ€๋ฅผ ์ง์ ‘ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋Š” ์ง€์—ญ์ด๊ธฐ์— ์„ ์ •ํ•˜์˜€๋‹ค(Fig. 1).

Fig. 1.

Study Site Location

Figure_KSCE_41_03_14_F1.jpg

ํ•ญ๊ณต์ดฌ์˜์— ์•ž์„œ ํ˜„์žฅ์กฐ์‚ฌ๊ฐ€ ๊ฐ€๋Šฅํ•œ ์ง€์—ญ์„ ์„ ๋ณ„ํ•˜๊ณ  ๋ฌด์ธํ•ญ๊ณต๊ธฐ๋ฅผ ์ด์šฉํ•œ ํ•ญ๊ณต์ดฌ์˜์„ ์œ„ํ•œ ํ—ˆ๊ฐ€ ๋“ฑ ์‚ฌ์ „ ์ค€๋น„ ์ž‘์—…์ด ํ•„์š”ํ•˜๋‹ค. ์ฆ‰, ๋Œ€์ƒ์ง€์˜ ๋†’์ด๋‚˜ ๋ฉด์ ์„ ๊ณ ๋ คํ•˜์—ฌ ๋ฌด์ธํ•ญ๊ณต๊ธฐ์˜ ์ข…๋ฅ˜๋ฅผ ์„ ํƒํ•˜์—ฌ์•ผ ํ•˜๋ฉฐ, ์ด์™€ ๋™์‹œ์— ๋น„ํ–‰๊ณ„ํš์„ ์ˆ˜๋ฆฝํ•˜์˜€๋‹ค. ์ด๋Š” ํ•ญ๊ณต์ดฌ์˜ ์žฅ๋น„ ์„ ํƒ์— ์žˆ์–ด์„œ๋„ ๋ชฉ์ ์— ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ์„ผ์„œ๋ฅผ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๊ณ  ์˜์ƒ ๋ฐ์ดํ„ฐ์˜ ํŠน์„ฑ์— ๋”ฐ๋ผ์„œ ํ”ผํ•ด๋ชฉ์˜ ์‹๋ณ„ ๋ฐฉ๋ฒ•์ด ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.

๋ฐ์ดํ„ฐ ๋ถ„์„์„ ์œ„ํ•ด์„œ๋Š” ์‚ฌ์ง„์ธก๋Ÿ‰ ํ”„๋กœ๊ทธ๋žจ๊ณผ GISํ”„๋กœ๊ทธ๋žจ์„ ํ™œ์šฉํ•œ ์œก์•ˆ๊ด€์ธก, ํ˜„์žฅ์กฐ์‚ฌ ๊ฒฐ๊ณผ, ํ•™์Šต๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ ์ˆœ์œผ๋กœ ์ง„ํ–‰๋˜์—ˆ์œผ๋ฉฐ, ํ”ผํ•ด๋ชฉ์˜ ์ž๋™ ํƒ์ง€๋ฅผ ์œ„ํ•ด์„œ ์‚ฌ์šฉํ•œ ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์ด๋ฏธ์ง€ ๋ถ„ํ• ๋ฐฉ๋ฒ•(SegNet)๊ณผ ์ด๋ฏธ์ง€ ๊ฒ€์ถœ๋ฐฉ๋ฒ•(YOLOv2)์˜ ๋‘ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ ์ž๋™ํƒ์ง€์— ์ ํ•ฉํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ด๋ณด์•˜๋‹ค.

3.2 ์˜์ƒ๋ฐ์ดํ„ฐ ํš๋“ ๋ฐ ์ •์‚ฌ์˜์ƒ ์ œ์ž‘

๋ฌด์ธํ•ญ๊ณต๊ธฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ•ญ๊ณต์ดฌ์˜์„ ์ง„ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์‚ฌ์ „์— ์ดฌ์˜ ์žฅ์†Œ ํ™•์ธ๊ณผ ์žฅ๋น„์˜ ์„ฑ๋Šฅ์„ ํŒŒ์•…ํ•ด์•ผ ํ•œ๋‹ค. ํŠนํžˆ, ๋ถ„์„์— ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์˜์ƒ ํ•ด์ƒ๋„๋Š” ์ •ํ™•๋„์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋ฉฐ, ์ด๋ฅผ ์„ค์ •ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ง€์ƒํ‘œ๋ณธ๊ฑฐ๋ฆฌ ๋˜๋Š” ์ง€์ƒํ•ด์ƒ๋ ฅ์„ ๋œปํ•˜๋Š” GSD (Ground Sampling Distance) ๊ณ„์‚ฐ์„ ์œ„ํ•ด์„œ ์ดฌ์˜๊ณ ๋„, ์ค‘๋ณต๋„, ์นด๋ฉ”๋ผ ํ•ด์ƒ๋„๋ฅผ ์ถฉ๋ถ„ํžˆ ์‚ฌ์ „์— ๊ฒ€ํ† ํ•ด์•ผ ํ•œ๋‹ค(Do et al., 2018).

์—ฐ๊ตฌ์— ํ™œ์šฉํ•œ ํ•ญ๊ณต์ดฌ์˜ ์žฅ๋น„๋กœ๋Š” ๋ฌด์ธํ•ญ๊ณต๊ธฐ์™€ ์นด๋ฉ”๋ผ๊ฐ€ ์žˆ๋‹ค. ๋จผ์ €, ์‚ฐ๋ฆผ์ง€์—ญ๊ณผ ๊ฐ™์ด ๋„“์€ ์ง€์—ญ์˜ ์ดฌ์˜๊ณผ ๋งคํ•‘์„ ์œ„ํ•ด์„œ๋Š” ๊ณ ์ •์ต ๋ฌด์ธํ•ญ๊ณต๊ธฐ๊ฐ€ ํ•„์š”ํ•˜์˜€์œผ๋ฉฐ, ๊ทธ์ค‘ KD-2๋Š” 40~60 km/h์˜ ์†๋„๋กœ ๋„“์€ ์ง€์—ญ์„ ๋น„ํ–‰ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋น„ํ–‰์‹œ๊ฐ„์€ ์ตœ๋Œ€ 50~60๋ถ„๊นŒ์ง€ ๊ฐ€๋Šฅํ•œ ๊ธฐ์ข…์ด๋‹ค. ๋ฐ์ดํ„ฐ ์ทจ๋“์„ ์œ„ํ•˜์—ฌ ์‚ฌ์šฉํ•œ ์นด๋ฉ”๋ผ๋กœ Sony ์‚ฌ์˜ RX1R2์ด๋ฉฐ ํ•ด์ƒ๋„๊ฐ€ 4,240๋งŒ ํ™”์†Œ๋กœ ๊ณ ์„ฑ๋Šฅ์˜ DSLR ์นด๋ฉ”๋ผ๋ฅผ ์ด์šฉํ•˜์˜€๋‹ค(Table 1).

Table 1.

Aerial Photography Equipment Specifications

Figure_KSCE_41_03_14_T1-1.jpg
(a) KD-2
Figure_KSCE_41_03_14_T1-2.jpg
(b) Sony RX1R2
Length 180 ร— 110 cm Weight 480 g
Weight 2.9 kg Sensor 35 mm F2.0 Zeiss Sonnar T lens with macro capability
Mounting weight 550 g Pixels 42.4 MP
Flight time 50โ€“60 min ISO sensitivity ISO 100-25600
Flight speed 40โ€“60 km/h Focal length f=35 mm / F2
Altitude 1000 m Focus range 24 cm-โˆž (Normal mode), 14โ€“29 cm (Macro mode)

ํ•ญ๊ณต์ดฌ์˜์€ 2019๋…„ 4์›”์— ์ฃผ๋ณ€ ์ง€ํ˜•๊ณผ ๋Œ€์ƒ์ง€์˜ ๋†’์ด, ๊ธฐ์ƒ์ƒํƒœ, ๋ฐฐํ„ฐ๋ฆฌ ์„ฑ๋Šฅ ๋“ฑ์„ ๊ณ ๋ คํ•˜์—ฌ ์ด๋ฅ™ ์ง€์ ์œผ๋กœ๋ถ€ํ„ฐ ๋น„ํ–‰๊ณ ๋„ 350 m, ์ข…โ€งํšก์ค‘๋ณต๋ฅ  80 %๋กœ ์„ค์ •ํ•˜์—ฌ ์ด๋ฃจ์–ด์กŒ์œผ๋ฉฐ, ํŠนํžˆ ๋น„ํ–‰๊ฒฝ๋กœ๋Š” ์ดฌ์˜์ง€์—ญ๋ณด๋‹ค ๊ธธ๊ฒŒ ์ฃผ์–ด์„œ ๊ณ ์ •์ต ๋ฌด์ธํ•ญ๊ณต๊ธฐ์˜ ๋น„ํ–‰ ํŠน์„ฑ์— ๋งž๊ฒŒ ์ถฉ๋ถ„ํ•œ ํšŒ์ „๋ฐ˜๊ฒฝ์„ ์„ค์ •ํ•˜์˜€๋‹ค.

์ดฌ์˜ํ•œ ์˜์ƒ ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ ๋ถ„์„ ๋Œ€์ƒ ์ง€์—ญ์˜ ์œ„์น˜๋‚˜ ์ขŒํ‘œ์™€ ๊ฐ™์€ ์ •๋ณด๋ฅผ ํš๋“ํ•˜๊ธฐ์—๋Š” ์–ด๋ ค์›€์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์—ฌ๋Ÿฌ ์žฅ์˜ ์‚ฌ์ง„์„ ์ง€๋„๋กœ ์ œ์ž‘ํ•˜๋Š” ๊ณผ์ •์ธ ์‚ฌ์ง„์ •ํ•ฉ ๊ณผ์ •์ด ํ•„์š”ํ•˜๋‹ค. ๋˜ํ•œ, ํ•ญ๊ณต์‚ฌ์ง„ ์ดฌ์˜์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๊ธฐํ•˜ํ•™์  ์™œ๊ณก๊ณผ ๊ฒฝ์‚ฌ์™œ๊ณก์„ ๋ณด์ •ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์˜์ƒ ๋ฐ์ดํ„ฐ์˜ ์œ„๋„, ๊ฒฝ๋„, ๊ณ ๋„๋ฅผ ํฌํ•จํ•œ 3์ฐจ์› GPS์ •๋ณด๊ฐ€ ์žˆ์–ด์•ผ ํ•œ๋‹ค. ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์ง€ ์•Š์€ ๊ฒฝ์šฐ์—๋Š” Geo-tagging ์ž‘์—…์„ ํ†ตํ•ด ๋ฌด์ธํ•ญ๊ณต๊ธฐ์— ์žฅ์ฐฉ๋˜์–ด ์žˆ๋Š” GPS ์ˆ˜์‹ ๊ธฐ์— ๊ธฐ๋ก๋œ ๋น„ํ–‰๋กœ๊ทธ ์ •๋ณด๋ฅผ ์ด์šฉํ•˜์—ฌ ์˜์ƒ ๋ฐ์ดํ„ฐ์— ์ž…๋ ฅํ•ด ์ฃผ์–ด์•ผ ํ•œ๋‹ค.

์ •์‚ฌ์˜์ƒ ์ œ์ž‘์„ ์œ„ํ•ด ์‚ฌ์šฉํ•œ ์†Œํ”„ํŠธ์›จ์–ด๋Š” Pix4D Mapper๋กœ ์›ํ•˜๋Š” ๋Œ€์ƒ์„ 3์ฐจ์› ๋ชจ๋ธ๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋ฐ์ดํ„ฐ ์ž…๋ ฅ ์‹œ ๋‹ค์–‘ํ•œ ํฌ๋งท์˜ ์‚ฌ์ง„์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ดฌ์˜์— ์‚ฌ์šฉํ•œ ์นด๋ฉ”๋ผ๊ฐ€ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ์ดˆ์  ๊ธธ์ด์™€ ์„ผ์„œ ํฌ๊ธฐ, ๋‚ด๋ถ€ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฐ’, ๋ Œ์ฆˆ ์™œ๊ณก ๋“ฑ์˜ ์ •๋ณด๋ฅผ ์ตœ์ ํ™”ํ•˜์—ฌ ๊ฐ์ฒด ์ƒ์„ฑ์— ๋ฐ˜์˜ํ•  ์ˆ˜ ์žˆ๋Š” ์žฅ์ ์ด ์žˆ๋‹ค. ํŠนํžˆ, ๋ถ„์„์„ ์œ„ํ•˜์—ฌ ๊ฐ์ฒด์˜ ๋ฐ˜์‚ฌ๋„ ๋˜๋Š” ์‹์ƒ์ง€์ˆ˜ ๋“ฑ์˜ ์ธ๋ฑ์Šค ๊ณ„์‚ฐ์„ ํ†ตํ•˜์—ฌ ํ‘œํ˜„๋˜๋Š” ์ •์‚ฌ์˜์ƒ์„ ๊ฐ ํ˜•์‹์˜ ๋งž๋Š” ๊ฒฐ๊ณผ๋ฌผ๋กœ ์ถœ๋ ฅํ•ด ์ฃผ๋Š” ๋“ฑ ๋‹ค์–‘ํ•œ ํ™œ์šฉ์ด ๊ฐ€๋Šฅํ•˜๋‹ค.

์‚ฌ์ง„ ์ •ํ•ฉ์— ํ™œ์šฉํ•œ ๋ฐ์ดํ„ฐ์˜ ์ˆ˜๋Š” 226์žฅ์œผ๋กœ Pix4D Mapper S/W๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ •์‚ฌ์˜์ƒ์„ ์ œ์ž‘ํ•˜๋Š” ๊ณผ์ •์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค. ๋จผ์ €, Geo-tagging ํ•˜์—ฌ GPS ์ขŒํ‘œ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ์˜์ƒ ๋ฐ์ดํ„ฐ๋ฅผ WGS84 (World Geodetic System 1984) ์ขŒํ‘œ๊ณ„๋กœ ์ž…๋ ฅํ•˜๊ฒŒ ๋˜๋ฉด ์‚ฌ์ง„์˜ ์œ„์น˜, ์ดฌ์˜ ๋ฐฉํ–ฅ ๋“ฑ์ด ๋‚˜ํƒ€๋‚˜๊ฒŒ ๋œ๋‹ค. ๋˜ํ•œ, ๋ฐ์ดํ„ฐ์— ๊ธฐ์ดˆํ•˜์—ฌ ๋งŒ๋“ค์–ด์ง€๋Š” ์ถœ๋ ฅ๋ฌผ์˜ GPS ์ขŒํ‘œ๊ณ„๋Š” ์ดฌ์˜์ง€์—ญ์„ ๊ธฐ์ค€์œผ๋กœ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํ•œ๊ตญ ๊ธฐ์ค€์˜ Korea 2000 ๋™๋ถ€ ์›์ ์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ๋ถ„์„์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค.

์ดˆ๊ธฐ ์ฒ˜๋ฆฌ ๊ณผ์ •์—์„œ๋Š” ์—ฌ๋Ÿฌ ์žฅ์˜ ์ด๋ฏธ์ง€๋ฅผ ๋งค์นญํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ํ‚คํฌ์ธํŠธ ์ด๋ฏธ์ง€ ์ˆ˜๋ฅผ ์„ค์ •ํ•˜๊ณ  ์นด๋ฉ”๋ผ ๋‚ด๋ถ€ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฐ’์„ ์ตœ์ ํ™”ํ•˜์—ฌ ๊ธฐ์ดˆ ๋ถ„์„ ์ž‘์—…์„ ์ง„ํ–‰ํ•˜๊ฒŒ ๋˜๋ฉฐ ๋ฐ์ดํ„ฐ ๋ถ„์„ ๊ฒฐ๊ณผ๋ฅผ ์ ๊ฒ€ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ’ˆ์งˆ ๋ณด๊ณ ์„œ๊ฐ€ ์ƒ์„ฑ๋œ๋‹ค.

๋‘ ๋ฒˆ์งธ ๊ณผ์ •์—์„œ๋Š” ๋Œ€์ƒ์„ 3D ๊ณต๊ฐ„์œผ๋กœ ๋‚˜ํƒ€๋‚ด์ฃผ๋Š” ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ(point cloud)๋ฅผ ์ƒ์„ฑํ•˜๊ฒŒ ๋˜์–ด ์ „์ฒด์ ์ธ ์ •์‚ฌ์˜์ƒ ์ด์ „์˜ ๊ฐ์ฒด ํ˜•ํƒœ๊ฐ€ ๋‚˜ํƒ€๋‚˜๊ณ  ๊ฐ ํฌ์ธํŠธ์—์„œ GPS ์ •๋ณด๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ์ถ”๊ฐ€ ๊ธฐ๋Šฅ์œผ๋กœ ๊ฐ ํฌ์ธํŠธ๋ฅผ ์ด์–ด ๋ฉด์„ ์ƒ์„ฑํ•˜๋Š” ์ž‘์—…์ธ ์‚ผ๊ฐ๋งค์‰ฌ ๊ธฐ๋Šฅ์„ ํ†ตํ•ด์„œ ๋Œ€์ƒ์˜ ํ˜•ํƒœ๋ฅผ ์ข€ ๋” ๊ตฌ์ฒดํ™”ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋งˆ์ง€๋ง‰ ๊ณผ์ •์œผ๋กœ ์˜์ƒ ๋ฐ์ดํ„ฐ๊ฐ€ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ์ธ๋ฑ์Šค ๊ฐ’์„ ์ž…๋ ฅํ•˜๋ฉด ๋‹ค์–‘ํ•œ ํ˜•ํƒœ์˜ ๋ถ„์„ ์˜์ƒ์„ ์–ป์„ ์ˆ˜ ์žˆ๋‹ค.

4. ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜์˜ ๊ฐ์ฒด ์ธ์‹๊ณผ ํŒ๋ณ„

๋ณธ ์ ˆ์—์„œ๋Š” ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ์˜ ์ธ์‹๊ณผ ํŒ๋ณ„์˜ ํšจ์œจ์„ฑ์„ ๊ณ ๋ คํ•˜์—ฌ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์–‘์ด ์ ์€ ๊ฒฝ์šฐ ์ƒ๋Œ€์ ์œผ๋กœ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜์˜ ๊ฐ์ฒด ์ธ์‹ ๋ฐฉ๋ฒ•๋ก ์„ ์ œ์‹œํ•˜๊ณ ์ž ํ•œ๋‹ค.

4.1 SegNet๊ณผ YOLO ๊ธฐ๋ฐ˜ ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜

์˜์ƒ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ์ด์šฉํ•˜์—ฌ ๊ฐ์ฒด๋ฅผ ํ•™์Šต์‹œํ‚ค๊ณ  ํƒ์ง€ํ•˜๋Š” ๊ธฐ์ˆ ์€ ํฌ๊ฒŒ ์ด๋ฏธ์ง€ ๋ถ„ํ• ๋ฐฉ๋ฒ•(Image Segmentation)๊ณผ ์ด๋ฏธ์ง€ ๊ฒ€์ถœ๋ฐฉ๋ฒ•(Image Detection)์ด ์žˆ๋‹ค. ์ด๋ฏธ์ง€ ๋ถ„ํ• ๋ฐฉ๋ฒ• ๊ฐ€์šด๋ฐ ๋Œ€ํ‘œ์ ์ธ ๊ธฐ์ˆ ๋กœ์จ ์ž์œจ์ฃผํ–‰ ์—ฐ๊ตฌ ๋ถ„์•ผ์— ๋งŽ์ด ํ™œ์šฉ๋˜๋Š” SegNet ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ํ”ฝ์…€(Pixel) ๋‹จ์œ„๋กœ ๊ฐ์ฒด๋ฅผ ์ธ์‹ํ•˜๊ณ  ํ•ด๋‹น ์œ„์น˜์˜ ํ”ฝ์…€์ด ์–ด๋–ค ๋ฌผ์ฒด๋ฅผ ์˜๋ฏธํ•˜๋Š”์ง€๋ฅผ ์˜ˆ์ธกํ•˜๋Š” Semantic Segmentation ๊ฐœ๋…์— ๊ธฐ์ดˆํ•˜๊ณ  ์žˆ๋‹ค. Fig. 2์—์„œ ๋ณด๋Š” ๋ฐ”์™€ ๊ฐ™์ด ์ด ๋ฐฉ๋ฒ•์€ ํ•™์Šต์˜ ์†๋„๋ฅผ ๋†’์ด๊ณ , ์œ„์น˜์ •๋ณด ์†์‹ค์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๋„คํŠธ์›Œํฌ ๊ตฌ์กฐ์ƒ convolution๊ณผ maxpooling์„ ํ†ตํ•ด ์••์ถ•ํ•˜๊ณ , ๋‹ค์‹œ upsampling ํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค(Badrinarayanan et al., 2017).

Fig. 2.

SegNet Architecture (Badrinarayanan et al., 2017)

Figure_KSCE_41_03_14_F2.jpg

์ด๋ฏธ์ง€ ๊ฒ€์ถœ๋ฐฉ๋ฒ•(Image Detection)์˜ ๋Œ€ํ‘œ์ ์ธ ๊ธฐ์ˆ ๋กœ๋Š” ์‹ค์‹œ๊ฐ„ ๊ฒ€์ถœ ๋ฐฉ๋ฒ•์ธ YOLOv2 ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ™œ์šฉํ•˜์˜€๋‹ค. YOLO ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๊ธฐ์กด ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ์˜ ๋Š๋ฆฐ ์†๋„์™€ ๋‚ฎ์€ ์ •ํ™•๋„ ๋ฌธ์ œ๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด์„œ CNN (Convolutional Neural Network) ๊ตฌ์กฐ๋ฅผ ๊ฐ€์ง€๊ณ  ํ•œ ์žฅ์˜ ์ด๋ฏธ์ง€์— ๋Œ€ํ•˜์—ฌ ๊ฒฝ๊ณ„ ๋ฐ•์Šค์˜ ์œ„์น˜์™€ ํด๋ž˜์Šค์˜ ๋ถ„๋ฅ˜๊ฐ€ ๋™์‹œ์— ์ด๋ฃจ์–ด์ง€๋Š” ์žฅ์ ์ด ์žˆ๋‹ค(Redmon et al., 2016). ๋‚˜์•„๊ฐ€ ์†๋„ ๊ฐœ์„ ์„ ์œ„ํ•˜์—ฌ ๋„คํŠธ์›Œํฌ ๊ตฌ์กฐ์—์„œ 2๊ฐœ์˜ FC (Fully Connected) layer๋ฅผ ์ œ๊ฑฐํ•˜์˜€๊ณ , ์•ต์ปค ๋ฐ•์Šค๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด์„œ ์„ฑ๋Šฅ ๋ถ€๋ถ„์—์„œ ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์–ด ์ด๋ฏธ์ง€ ์‚ฌ์ด์ฆˆ์— ๋”ฐ๋ผ ๊ฒ€์ถœ ์†๋„์™€ ์ •ํ™•๋„๋ฅผ ์ปจํŠธ๋กคํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์ด YOLOv2์˜ ํฐ ํŠน์ง•์ด๋‹ค(Fig. 3).

Fig. 3.

YOLO Architecture (Giuffrida et al., 2019)

Figure_KSCE_41_03_14_F3.jpg

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ ๊ฐ์ฒด ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ๋น ๋ฅด๊ณ  ์ •ํ™•ํ•˜๊ฒŒ ๊ฒ€์ถœํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ์ด๋ฏธ์ง€ ๋ถ„ํ• ๋ฐฉ์‹์˜ ๋Œ€ํ‘œ์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ SegNet๊ณผ ์ด๋ฏธ์ง€ ๊ฒ€์ถœ๋ฐฉ์‹์˜ ๋Œ€ํ‘œ์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ YOLOv2์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ด๋ณด์•˜๋‹ค.

4.2 ํ•™์Šต ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ๊ณผ ํŒŒ๋ผ๋ฉ”ํƒ€ ์„ค์ •

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ฌด์ธํ•ญ๊ณต๊ธฐ์—์„œ ์ดฌ์˜ํ•œ ์˜์ƒ๊ธฐ๋ฐ˜์˜ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ ํŒ๋ณ„ ๊ณผ์ •๊ณผ ๋ณ„๋„๋กœ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ์˜ ์ฐธ๊ฐ’์„ ํ™•๋ณดํ•˜๊ธฐ ์œ„ํ•œ ํ˜„์žฅ์กฐ์‚ฌ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•˜๊ธฐ๋กœ ํ•œ๋‹ค.

๋จผ์ €, ์•ž ์ ˆ์—์„œ ์ œ์ž‘๋œ ์ •์‚ฌ์˜์ƒ์—์„œ ArcGIS S/W๋ฅผ ํ™œ์šฉํ•˜์—ฌ GPS ์œ„์น˜์ •๋ณด๋ฅผ ํš๋“ํ•˜์˜€๋‹ค. ์ด ์ขŒํ‘œ ์ •๋ณด๋Š” ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด๋ชฉ์˜ ํŠน์ง•์  ์ถ”์ถœ๊ณผ ๋ผ๋ฒจ๋ง(labeling) ๊ณผ์ •์„ ํ†ตํ•ด ํ•™์Šต ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•œ๋‹ค. ์ •์‚ฌ์˜์ƒ์—์„œ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ์˜์‹ฌ๋ชฉ์œผ๋กœ ์ถ”์ •๋˜๋Š” ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ์œผ๊ธฐ ์œ„ํ•ด ์ฃผ๋ณ€ ๊ฑด๊ฐ•ํ•œ ๋‚˜๋ฌด๋“ค๊ณผ ๋‹ค๋ฅด๊ฒŒ ์žŽ์ด ๋ณ€์ƒ‰์ด ๋˜๊ฑฐ๋‚˜ ๋–จ์–ด์ ธ ๊ฐ€์ง€๋งŒ ๋‚จ์€ ๋‚˜๋ฌด๋ฅผ ์„ ๋ณ„ํ•˜๊ณ  ์ขŒํ‘œ ์ •๋ณด๋ฅผ ๋งค์นญํ•˜๋Š” ๊ณผ์ •์„ ์ˆ˜ํ–‰ํ•˜์—ฌ, 198๊ฐœ์˜ ์ƒ˜ํ”Œ์„ ์–ป์—ˆ๋‹ค.

Fig. 4๋Š” 198๊ฐœ์˜ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ์˜ ์œ„์น˜๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์œผ๋กœ ์‹ค์ œ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘์— ๊ฐ์—ผ๋œ ๊ฒƒ์„ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ์ดˆ ์ •๋ณด์ด๋‹ค. ์ •์‚ฌ์˜์ƒ์—์„œ ์ถ”์ •ํ•œ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ์ด ์‹ค์ œ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘์— ๊ฐ์—ผ๋œ ๊ฒƒ์ธ์ง€๋ฅผ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•ด ํ˜„์žฅ์กฐ์‚ฌ๋ฅผ ์‹ค์‹œํ•˜์˜€๋‹ค(Fig. 5). ์ด ๊ณผ์ •์€ ๊ธฐ์กด ์—ฐ๊ตฌ์™€์˜ ์ฐจ๋ณ„์„ฑ์œผ๋กœ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด๋ชฉ์˜ ํŠน์ง•์  ์ถ”์ถœ์„ ํ†ตํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ ๊ตฌ์ถ•๊ณผ ์ •์‚ฌ์˜์ƒ์—์„œ ์ถ”์ •ํ•œ ํ”ผํ•ด๋ชฉ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋ฅผ ์‹ค์ œ๋กœ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ๊ผญ ํ•„์š”ํ•œ ๊ณผ์ •์ด๋‹ค.

Fig. 4.

Observation Results of Pine Wilt Disease Based on Image

Figure_KSCE_41_03_14_F4.jpg

Fig. 5.

Field Survey

Figure_KSCE_41_03_14_F5.jpg

์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ ํ‰๊ฐ€๋ฅผ ๋ชฉ์ ์œผ๋กœ ํ•œ ํ˜„์žฅ์กฐ์‚ฌ์—์„œ ์‹ค์ œ ๊ฐ์—ผ๋ชฉ์˜ ์œ„์น˜(์ขŒํ‘œ) ํ™•์ธ์„ ์˜์ƒ ์ž๋ฃŒ๋กœ ์ถ”์ •ํ•œ 198๊ฐœ์˜ ์ •๋ณด์™€ ์ผ๋Œ€์ผ๋กœ ํ™•์ธํ•˜์˜€๋‹ค. ๋ถ„์„ ๋Œ€์ƒ์ง€๋Š” ํ•ด๋‹น ์ง€์ž์ฒด์—์„œ ์ด๋ฏธ ๊ฐ์—ผ๋œ ์†Œ๋‚˜๋ฌด๋“ค์„ ๋ถ‰์€์ƒ‰์œผ๋กœ ๋งˆํ‚นํ•˜๊ณ  ๋ผ๋ฒจ์ง€๋ฅผ ๋ถ™์—ฌ ํ‘œ์‹œํ•ด ๋†“์€ ์ง€์—ญ์œผ๋กœ ์ด๋ฅผ ๊ธฐ์ค€์œผ๋กœ GPS ์žฅ๋น„์™€ ์•ผ์žฅ์œผ๋กœ ๊ธฐ์กด ๊ฐ์—ผ๋ชฉ๊ณผ ์ƒˆ๋กœ ๋ฐœ์ƒํ•œ ๊ฐ์—ผ๋ชฉ์˜ ์œ„์น˜๋ฅผ ํŒŒ์•…ํ•˜์˜€๋‹ค.

ํ˜„์žฅ์กฐ์‚ฌ๋Š” 2019๋…„ 4์›”์— ์˜ค์ „ 10์‹œ๋ถ€ํ„ฐ ํ•ด๊ฐ€ ์งˆ ๋ฌด๋ ต๊นŒ์ง€ 2๋ช…์”ฉ 3๊ฐœ ์กฐ๋กœ 2ํšŒ ์‹ค์‹œํ•˜์˜€์œผ๋ฉฐ, ์ตœ๋Œ€ํ•œ ๋งŽ์€ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ์˜ ์œ„์น˜์™€ ํŠน์ง•์ ์„ ํš๋“ํ•˜๊ธฐ ์œ„ํ•ด ์ ‘๊ทผ์ด ๋ถˆ๊ฐ€๋Šฅํ•œ ๊ธ‰๊ฒฝ์‚ฌ์ง€๋‚˜ ์ ˆ๋ฒฝ๊ณผ ๊ฐ™์€ ์œ„ํ—˜ํ•œ ์ง€ํ˜• ์ด์™ธ์— ์ „์ˆ˜์กฐ์‚ฌ๋ฅผ ์‹ค์‹œํ•œ ๊ฒฐ๊ณผ, ๊ด€์ธกํ•œ 198๊ฐœ ์ค‘์— 84๊ฐœ์˜ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ์˜ ์ฐธ๊ฐ’์„ ํš๋“ํ•˜์˜€๋‹ค.

ํ˜„์žฅ์กฐ์‚ฌ๋ฅผ ํ†ตํ•ด ์–ป์€ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ์˜ ์œ„์น˜์ •๋ณด๋ฅผ ์ด์šฉํ•˜์—ฌ ๋ฌด์ธํ•ญ๊ณต๊ธฐ ์˜์ƒ ์ž๋ฃŒ์—์„œ ์–ป์€ ํ•ญ๊ณต์˜์ƒ์˜ ๊ด€์ธก์ ๋“ค๊ณผ ๋งค์นญํ•  ์ˆ˜ ์žˆ์—ˆ์œผ๋ฉฐ, ๋”ฅ๋Ÿฌ๋‹ ํ•™์Šต์„ ์œ„ํ•œ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ์˜ ์ด๋ฏธ์ง€๋ฅผ ์ถ”์ถœํ•˜์˜€๋‹ค. ์ด๋•Œ Fig. 6์—์„œ ๋ณด๋Š” ๋ฐ”์™€ ๊ฐ™์ด ์ •์‚ฌ์˜์ƒ์„ 50 m ร— 50 m ํฌ๊ธฐ์˜ ๊ทธ๋ฆฌ๋“œ๋กœ ๋‚˜๋ˆ„๊ณ  ๊ฐ ๊ตฌ์—ญ์— ๋Œ€ํ•ด์„œ ๋™์ผํ•œ ํ”ฝ์…€์˜ ํฌ๊ธฐ๋กœ ์ด๋ฏธ์ง€๋ฅผ ์ถ”์ถœํ•˜์˜€๋‹ค. ๋˜ํ•œ, ์ถฉ๋ถ„ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ํ™•๋ณดํ•˜๊ธฐ ์œ„ํ•ด ํ•˜๋‚˜์˜ ์ด๋ฏธ์ง€์— ๋Œ€ํ•˜์—ฌ 90ยฐ, 180ยฐ, 270ยฐ๋งŒํผ ํšŒ์ „์‹œ์ผœ ์ด 1,425์žฅ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ™•๋ณดํ•˜์˜€๋‹ค.

Fig. 6.

Learning Data Set

Figure_KSCE_41_03_14_F6.jpg

๊ตฌ์ถ•ํ•œ ๋ฐ์ดํ„ฐ ๊ฐ€์šด๋ฐ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ์„ ์ค‘์‹ฌ์œผ๋กœ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•˜์˜€์œผ๋ฉฐ, ๊ฐ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ํ•ด๋‹นํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ ๋ผ๋ฒจ๋ง ์ž‘์—…์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค. ๋จผ์ €, SegNet์€ ์ด๋ฏธ์ง€์— ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด๋ชฉ์ด ์žˆ๋Š” ๊ณณ์„ ๊ฒ€์€์ƒ‰, ๊ทธ ์™ธ ๋ถ€๋ถ„์„ ํฐ์ƒ‰์œผ๋กœ ๋ผ๋ฒจ๋ง ํ•˜์˜€๊ณ , YOLOv2๋Š” Bounding box ํ˜•ํƒœ์˜ ๋ผ๋ฒจ๋ง์œผ๋กœ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด๋ชฉ๊ณผ ์•„๋‹Œ ๊ฒƒ์œผ๋กœ ๊ตฌ๋ถ„ํ•˜์—ฌ ์ž‘์—…์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค. ์—ฌ๊ธฐ์„œ ๋”ฅ๋Ÿฌ๋‹ ํ•™์Šต์— ์˜ค๋ฅ˜๋กœ ๋‚˜ํƒ€๋‚˜๋Š” ์ด๋ฏธ์ง€์˜ ๊ณต๋ฐฑ ๋ถ€๋ถ„๊ณผ ๊ทธ๋ฆผ์ž ๋“ฑ์˜ ๋ฐ์ดํ„ฐ๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ์—์„œ ์ œ์™ธํ•˜์˜€๋‹ค.

Table 2์—์„œ ๋ณด๋Š” ๋ฐ”์™€ ๊ฐ™์ด ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ ๊ฒฐ๊ณผ, SegNet๊ณผ YOLOv2 ๊ฐ๊ฐ 475์žฅ, 488์žฅ์œผ๋กœ ํ•™์Šต ๋ฐ์ดํ„ฐ ์–‘์€ ํฌ๊ฒŒ ์ฐจ์ด๊ฐ€ ์—†์—ˆ์œผ๋ฉฐ, 70 %๋ฅผ ํ•™์Šต์šฉ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉํ•˜๊ณ  ๋‚˜๋จธ์ง€ 30 %๋ฅผ ๊ฒ€์ฆ์šฉ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ๋ถ„ํ•˜์˜€๋‹ค.

Table 2.

Configuration Results of Deep Learning Data (Unit: number)

Data for labeling Data for learning Data for verification
SegNet YOLOv2 SegNet YOLOv2 SegNet YOLOv2
475 488 335 342 140 146

์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ ์—ฌ๋ถ€์— ๋”ฐ๋ฅธ ๋ผ๋ฒจ๋ง ์ž‘์—…์„ ๊ฑฐ์ณ ๋”ฅ๋Ÿฌ๋‹์„ ์œ„ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฐ’์„ ์„ค์ •ํ•ด์•ผ ํ•˜๋ฉฐ, ๊ธฐ์กด ๋‘ ๋ชจ๋ธ์˜ ๋ ˆ์ด์–ด๋ฅผ ํ™œ์šฉํ•˜๋˜ ๋งˆ์ง€๋ง‰ ๊ฐ์ฒด๋ฅผ ์ธ์‹ํ•˜๊ฑฐ๋‚˜ ์˜๋ฏธ๋ก ์ ์œผ๋กœ ๋ถ„ํ• ํ•˜๋Š” ๊ณณ์„ ๋ณ€๊ฒฝํ•˜์˜€๋‹ค(Fig. 7).

Fig. 7.

Learning Data Labeling

Figure_KSCE_41_03_14_F7.jpg

SegNet ๊ธฐ๋ฐ˜์˜ ํ•™์Šต์„ ์œ„ํ•˜์—ฌ ํ™œ์„ฑํ™” ํ•จ์ˆ˜๋Š” Sigmoid๋ฅผ ํ™œ์šฉํ•˜๊ณ  ํŠธ๋ ˆ์ด๋„ˆ๋Š” Adam์„ ํ™œ์šฉํ•˜์—ฌ ์ด 2,000 Epoch์˜ ํ•™์Šต์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค. ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ์˜ ์ž๋™ ๊ฐ์ฒด ์ธ์‹๊ณผ ํŒ๋ณ„ ๊ณผ์ •์„ ์‹คํ–‰ํ•œ ๊ฒฐ๊ณผ, ํ•˜๋‚˜์˜ ๋ฐ์ดํ„ฐ ๊ฒฐ๊ณผ์— ๋Œ€ํ•˜์—ฌ ์›๋ณธ, ๋ผ๋ฒจ๋ง, ๊ฒฐ๊ณผ ์ด๋ฏธ์ง€๊ฐ€ ์ถ”์ถœ๋˜์—ˆ์œผ๋ฉฐ, ๊ฐ๊ฐ์˜ ์ด๋ฏธ์ง€๋ฅผ ๋น„๊ตํ•˜๋ฉด ์ฐจ์ด๋ฅผ ํŒ๋ณ„ํ•  ์ˆ˜ ์žˆ๋‹ค.

YOLOv2์˜ ํ•™์Šต์„ ์œ„ํ•œ ํ™œ์„ฑํ™” ํ•จ์ˆ˜, ํŠธ๋ ˆ์ด๋„ˆ, Epoch์˜ ๊ฐ’์€ ๋ชจ๋‘ ๋™์ผํ•˜๋ฉฐ, ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ์„ ํ•˜๋‚˜์˜ ํด๋ž˜์Šค๋กœ ๋ณ€๊ฒฝํ•˜๋ฉด์„œ output layer์˜ ๊ฐ’์„ ์ค„์˜€๊ณ , input์˜ ํฌ๊ธฐ๋ฅผ ์ค€๋น„ํ•œ ๋ฐ์ดํ„ฐ์— ๋งž์ถ”์–ด ์ „์ฒด layer๋ฅผ ์ˆ˜์ •ํ•˜์˜€๋‹ค. ๊ฐ์ฒด์˜ ์ž๋™ ์ธ์‹์€ ํ•˜๋‚˜์˜ ์ด๋ฏธ์ง€ ์•ˆ์— Bounding box๋กœ ๋ผ๋ฒจ๋ง ํ•œ ๋ฐ์ดํ„ฐ์™€ ๊ฒ€์ถœ๋œ ๋ฐ์ดํ„ฐ๋กœ ๋‚˜ํƒ€๋‚˜๊ฒŒ ๋˜๋ฉฐ, ์ด๋•Œ ์ฐธ๊ฐ’์œผ๋กœ ์ œ์‹œํ•œ Bounding box๊ฐ€ ๋”ฅ๋Ÿฌ๋‹ ํ•™์Šต๊ฒฐ๊ณผ๋กœ ๋‚˜ํƒ€๋‚œ Bounding box์™€ ์ผ์ • ์˜์—ญ ์ค‘์ฒฉ๋˜์–ด์•ผ ์ฐธ๊ฐ’์œผ๋กœ ํŒ์ •๋˜๋ฉฐ, ์ด๋ฅผ IoU (Intersection over union)๋ผ๊ณ  ํ•œ๋‹ค. ์ด ๊ณผ์ •์—์„œ ๊ฐ™์€ ๊ฐ์ฒด์— ๋Œ€ํ•ด ๋งŽ์€ Bounding box๊ฐ€ ์ƒ์„ฑ๋˜๊ณ  ์ด ๊ฐ€์šด๋ฐ ์‹ ๋ขฐ์„ฑ์ด ๊ฐ€์žฅ ๋†’๊ณ  IoU๊ฐ€ ์ผ์ • ๊ฐ’ ์ด์ƒ์ธ Bounding box๋งŒ์„ ๋‚จ๊ธฐ๊ณ  ๋‚˜๋จธ์ง€๋Š” ์ œ๊ฑฐ๋œ๋‹ค.

๋˜ํ•œ, YOLOv2์˜ ๋ถ„์„์—์„œ ํŒŒ๋ผ๋ฏธํ„ฐ ๊ฐ’์ธ threshold์˜ ์กฐ์ ˆ์ด ๊ฐ€๋Šฅํ•˜๋‹ค. ์ด๋Š” ๊ธฐ์กด ํ•™์Šต์œผ๋กœ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•  ๊ฒฝ์šฐ threshold ๊ฐ’์— ๋”ฐ๋ผ ์ฃผ์–ด์ง„ ์ž„๊ณ„๊ฐ’์„ ๊ธฐ์ค€์œผ๋กœ ํ•ด๋‹น ๊ฐ์ฒด๋ฅผ ์ถœ๋ ฅํ•˜๊ฒŒ ๋œ๋‹ค. ๋”ฐ๋ผ์„œ ์ ์ ˆํ•œ threshold ๊ฐ’์„ ๋„์ถœํ•˜๊ธฐ ์œ„ํ•ด์„œ 0.1์—์„œ 0.9๊นŒ์ง€ ์กฐ์ ˆํ•˜์—ฌ ๋น„๊ต ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค.

4.3 ๋ถ„์„ ๊ฒฐ๊ณผ

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ์„ ํŒ๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด ์ฃผ๋กœ ํ™œ์šฉํ•˜๋Š” ๋‘ ๊ฐ€์ง€ ์ง€ํ‘œ์ธ ์ •๋ฐ€๋„(Precision)์™€ ์žฌํ˜„์œจ(Recall)์„ ์ด์šฉํ•˜์˜€๋‹ค. ์ •๋ฐ€๋„์™€ ์žฌํ˜„์œจ์„ ์‚ฐ์ •ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ํ”ผํ•ด๋ชฉ์„ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๊ฒ€์ถœํ•œ ๊ฒฐ๊ณผ(TP: True positive)์™€ ํ”ผํ•ด๋ชฉ์ด ์•„๋‹Œ ๊ฒƒ์„ ํ”ผํ•ด๋ชฉ์œผ๋กœ ์ž˜๋ชป ๊ฒ€์ถœํ•œ ๊ฒฐ๊ณผ(FP: False positive) ๊ทธ๋ฆฌ๋กœ ํ”ผํ•ด๋ชฉ์ด ์•„๋‹Œ ๊ฒƒ์„ ์•„๋‹Œ ๊ฒƒ์œผ๋กœ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๊ฒ€์ถœํ•œ ๊ฒฐ๊ณผ(FN: False negative)์˜ ๊ฐ’์„ ํ™œ์šฉํ•˜์—ฌ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ๋‹ค(Kotsiantis et al., 2006).

(1)
P r e c i s i o n = T P T P + F P = T P a l l   d e t e c t i o n s
(2)
R e c a l l = T P T P + F N = T P a l l   g r o u n d   t r u t h s

Table 3์˜ ๊ฒฐ๊ณผ์—์„œ ๋‘ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ TP, TF, FN ๊ฒ€์ถœ ๊ฐœ์ˆ˜์˜ ํ™•์—ฐํ•œ ์ฐจ์ด๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” ํ•™์Šต์„ ์œ„ํ•œ ๋ผ๋ฒจ๋ง ์ž‘์—… ์‹œ SegNet์€ ํ”ฝ์…€ ๊ฐœ์ˆ˜, YOLOv2๋Š” Bounding box ๊ฐœ์ˆ˜๊ฐ€ ํ•™์Šต๋˜๊ณ , ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ๊ฒฐ๊ณผ์—์„œ ํ”ฝ์…€๊ณผ Bounding box ๊ฐœ์ˆ˜๋กœ ๋‚˜ํƒ€๋‚˜๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.

Table 3.

Performance Comparison of Detection for Pine Wilt Disease (Unit: number)

Algorithm TP FP FN Precision Recall f1-score
SegNet 220,675 220,617 114,194 0.50 0.66 0.57
YOLOv2 (Threshold: 0.3) 204 62 58 0.77 0.78 0.77

ํ•œํŽธ ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•  ๋•Œ ์ •๋ฐ€๋„๋งŒ์œผ๋กœ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒฝ์šฐ ํด๋ž˜์Šค๋ณ„ ๋ฐ์ดํ„ฐ์˜ ๊ท ํ˜•์ด ๊ฑฐ์˜ ๋™์ผํ•  ๊ฒฝ์šฐ์—๋Š” ์ ์ ˆํ•˜๊ฒŒ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์œผ๋‚˜, ๋ฐ˜๋Œ€๋กœ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•œ์ชฝ์œผ๋กœ ์น˜์šฐ์น˜๊ฒŒ ๋˜๋ฉด ๋น„์ค‘์ด ํฐ ์ชฝ์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ๊ฒฐ๊ณผ์— ํฐ ์˜ํ–ฅ์„ ๋ผ์น˜๋Š” ๋ฌธ์ œ๊ฐ€ ์žˆ๋‹ค. ๋˜ํ•œ, ๋‘ ํ‰๊ฐ€ ์ง€ํ‘œ์˜ ๊ด€๊ณ„๋Š” ์ •๋ฐ€๋„๊ฐ€ ๋†’์œผ๋ฉด ์žฌํ˜„์œจ์ด ๋‚ฎ๊ณ , ์ •๋ฐ€๋„๊ฐ€ ๋‚ฎ์œผ๋ฉด ์žฌํ˜„์œจ์ด ๋†’์€ ๊ฒฝํ–ฅ์ด ๋‚˜ํƒ€๋‚˜๋Š” ๋‹จ์ ์„ ๋ณด์™„ํ•˜๊ธฐ ์œ„ํ•ด ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ƒˆ๋กœ์šด ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ๋„์ž…ํ•˜์˜€๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ ๋ฐ์ดํ„ฐ๋ณด๋‹ค ๋ฐ˜๋Œ€๋กœ ์ •์ƒ ์†Œ๋‚˜๋ฌด์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถˆ๊ท ํ˜•์ ์œผ๋กœ ๋งŽ๊ธฐ ๋•Œ๋ฌธ์— ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ตœ์ ํ™”๋œ ์„ฑ๋Šฅ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ๋ถˆ๊ท ํ˜•ํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด์„œ๋„ ์„ฑ๋Šฅ ์ธก์ •์ด ๊ฐ€๋Šฅํ•œ F-score๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค(Derczynski, 2016).

(3)
F - s c o r e = 2 P r e c i s i o n ร— R e c a l l ( P r e c i s i o n + R e c a l l )

f1-score ๊ฐ’์€ ์ •๋ฐ€๋„(Precision)์™€ ์žฌํ˜„์œจ(Recall)์˜ ์กฐํ™”ํ‰๊ท ์œผ๋กœ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ๋‹ค. Table 3์—์„œ ์•Œ ์ˆ˜ ์žˆ๋Š” ๋ฐ”์™€ ๊ฐ™์ด, SegNet ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ f1-score ๊ฐ’์ด 0.57๋กœ ๋‚˜ํƒ€๋‚ฌ๊ณ , YOLOv2 ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๊ฒฐ๊ณผ๋Š” threshold 0.3์—์„œ f1-score๊ฐ€ 0.77๋กœ ๋‚˜ํƒ€๋‚˜ YOLOv2๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ์šฐ์ˆ˜ํ•จ์„ ํ™•์ธํ•˜์˜€๋‹ค. ๋˜ํ•œ ๊ธฐ์กด์˜ ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜ ์˜์ƒ๋ถ„์„์„ ์œ„ํ•œ ์„ฑ๋Šฅ ํŒ๋ณ„ ์ง€ํ‘œ์ธ ์ •๋ฐ€๋„์™€ ์žฌํ˜„์œจ์„ ๋น„๊ตํ•˜๋”๋ผ๋„ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ ํƒ์ง€๋ฅผ ์œ„ํ•ด์„œ๋Š” YOLOv2๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ์šฐ์ˆ˜ํ•œ ๊ฒฐ๊ณผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค.

5. ๊ฒฐ ๋ก 

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ „๊ตญ์ ์œผ๋กœ ํ™•์‚ฐ๋˜๊ณ  ์žˆ๋Š” ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ๋ชฉ์˜ ํ”ผํ•ด ๊ทœ๋ชจ๋ฅผ ์ค„์ด๊ณ  ๋ฐฉ์ œ์— ์†Œ์š”๋˜๋Š” ๊ฒฝ์ œ์  ์†์‹ค์„ ์ ˆ๊ฐํ•˜๊ธฐ ์œ„ํ•ด ๋ฌด์ธํ•ญ๊ณต๊ธฐ์™€ ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•œ ์˜ˆ์ฐฐ ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•˜๊ณ  ํ™œ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ์‚ดํŽด๋ณด์•˜๋‹ค.

ํŠนํžˆ, ์šฐ๋ฆฌ๋‚˜๋ผ ์‚ฐ์ง€์˜ ํŠน์„ฑ์ƒ ์ ‘๊ทผ๋กœ๊ฐ€ ์—†๊ณ  ํ—˜์ค€ํ•œ ์ง€์—ญ์ด ๋งŽ๋‹ค๋Š” ์ ์„ ๊ฐ์•ˆํ•  ๋•Œ ์ธ๋ ฅ ์ค‘์‹ฌ์˜ ๊ธฐ์กด ์˜ˆ์ฐฐ ๋ฐฉ๋ฒ•์˜ ํ•œ๊ณ„๋ฅผ ๋“œ๋ก ์„ ํ™œ์šฉํ•œ ๋ฌด์ธํ•ญ๊ณต์ดฌ์˜์œผ๋กœ ์‹œ๊ฐ„๊ณผ ๋น„์šฉ์ ์ธ ์ธก๋ฉด์—์„œ ํ™œ์šฉ ๊ฐ€๋Šฅํ•จ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค.

๋˜ํ•œ, ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ํ•™์Šต์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ์˜ ์‹ ๋ขฐ์„ฑ์„ ํ™•๋ณดํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ๋Œ€์ƒ ์ง€์—ญ์˜ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ ์˜์‹ฌ๋ชฉ์„ ๋Œ€์ƒ์œผ๋กœ ์ •ํ•ฉํ•œ ์˜์ƒ๊ณผ ์‹ค์ œ ๊ฐ์—ผ ์—ฌ๋ถ€๋ฅผ ํ˜„์žฅ ์กฐ์‚ฌ๋ฅผ ํ†ตํ•ด ํ™•์ธํ•˜์˜€๋‹ค๋Š” ์ ์€ ๊ธฐ์กด ์—ฐ๊ตฌ ๋ฐฉ๋ฒ•๊ณผ์˜ ์ฐจ๋ณ„์„ฑ์ด๋ผ ํ•  ์ˆ˜ ์žˆ๋‹ค.

๋‚˜์•„๊ฐ€ ํ™•์ธ๋œ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด๋ชฉ ๋ฐ์ดํ„ฐ๋ฅผ ์ž๋™์œผ๋กœ ๊ฒ€์ถœํ•˜๊ธฐ ์œ„ํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ SegNet๊ณผ YOLOv2 ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ™œ์šฉํ•˜์˜€๋‹ค. ์ด๋Š” ํ˜„์žฌ๊นŒ์ง€ ์ ˆ๋Œ€์  ์šฐ์œ„๋ฅผ ์ ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ์—†๋Š” ํ˜„์‹ค์—์„œ ์˜์ƒ ๋ฐ์ดํ„ฐ ๊ธฐ๋ฐ˜์˜ ํ”ฝ์…€ ๋‹จ์œ„๋กœ ๊ฐ์ฒด์˜ ํ˜•ํƒœ์™€ ์ข…๋ฅ˜๋ฅผ ๋ถ„๋ฅ˜ํ•˜๋Š” SegNet๊ณผ ๋ฐ•์Šค ๋‹จ์œ„๋กœ ๊ฐ์ฒด์˜ ํŠน์ง•์„ ํ•™์Šตํ•˜์—ฌ ํƒ์ง€ํ•˜๋Š” YOLO์˜ ์ ‘๊ทผ ๋ฐฉ๋ฒ•์ด ๋‹ค๋ฅด์ง€๋งŒ, ์ผ์ •ํ•œ ๋ฉด์ ์œผ๋กœ ๊ฐ์—ผ ์—ฌ๋ถ€๋ฅผ ํ•™์Šต์‹œํ‚ค๊ณ  ์„ฑ๋Šฅ์„ ๋น„๊ตํ•œ๋‹ค๋Š” ์ ์—์„œ ์˜๋ฏธ๊ฐ€ ์žˆ๋‹ค๊ณ  ํŒ๋‹จ๋œ๋‹ค.

๋งˆ์ง€๋ง‰์œผ๋กœ ๊ฐ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์ •๋ฐ€๋„(Precision)์™€ ์žฌํ˜„์œจ(Recall)์„ ๊ฐ€์ง€๊ณ  F-score ๊ณ„์‚ฐํ•˜์˜€์œผ๋ฉฐ, ์ •ํ™•๋„๊ฐ€ ๊ฐ๊ฐ SegNet์ด 57 %, YOLOv2๊ฐ€ ์•ฝ 77 %๋กœ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด๋ชฉ์„ ๊ฒ€์ถœํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” YOLOv2๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ํ™•์ธํ•˜์˜€๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ ์ œ์‹œํ•œ ๋ฌด์ธํ•ญ๊ณต๊ธฐ๋ฅผ ์ด์šฉํ•œ ์˜์ƒ ๋ฐ์ดํ„ฐ์˜ ํš๋“ ๋ฐ ์ •์‚ฌ์˜์ƒ ์ œ์ž‘ ๊ณผ์ •, ๊ทธ๋ฆฌ๊ณ  ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜์˜ ์˜์ƒ๋ถ„์„์„ ํ†ตํ•œ ํ”ผํ•ด๋ชฉ ํƒ์ง€ ๋ฐฉ์•ˆ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ์„ฑ๊ณผ๋Š” ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ๊ฐ์—ผ์˜์‹ฌ๋ชฉ ์˜ˆ์ฐฐ๊ณผ ๋ฐฉ์ œ ์ž‘์—…์„ ์œ„ํ•œ ์‹ค๋ฌด์— ๋งŽ์€ ๋„์›€์ด ๋  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋œ๋‹ค.

ํ–ฅํ›„ ์—ฐ๊ตฌ๋กœ๋Š” ๋ฌด์ธํ•ญ๊ณต๊ธฐ๋ฅผ ์ด์šฉํ•œ ํ•ญ๊ณต์ดฌ์˜์˜ ๊ฒฝ์šฐ ๊ณ„์ ˆ ๋ฐ ์ผ์กฐ๋Ÿ‰์˜ ์ฐจ์ด์— ๋”ฐ๋ผ ์ •์‚ฌ์˜์ƒ ์ œ์ž‘์„ ์œ„ํ•œ ๋ณด์ •๊ฐ’์— ๋Œ€ํ•œ ๋ฉด๋ฐ€ํ•œ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•˜๋ฉฐ, ๋”ฅ๋Ÿฌ๋‹ ๋ถ„์„์— ๋ฌธ์ œ๊ฐ€ ๋˜์—ˆ๋˜ ์†Œ๋‚˜๋ฌด์žฌ์„ ์ถฉ๋ณ‘ ํ”ผํ•ด๋ชฉ๊ณผ ๋น„์Šทํ•œ ํŠน์ง•์ ์œผ๋กœ ํŒ๋ณ„๋˜๋Š” ์ง€๋ฉด, ๋ฌ˜์ง€, ๋งˆ๋ฅธํ’€ ๋“ฑ์˜ ์ถ”๊ฐ€ ํ•™์Šต์ž๋ฃŒ ๊ตฌ์ถ•์„ ํ†ตํ•ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ์„ ๋†’์ด๋Š” ๋ฐฉ์•ˆ๋„ ์ถ”๊ฐ€๋˜์–ด์•ผ ํ•  ๊ฒƒ์ด๋‹ค.

References

1 
Badrinarayanan, V., Kendall, A. and Cipolla, R. (2017). "Segnet: A deep convolutional encoder-decoder architecture for image segmentation." IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE, Vol. 39, No. 12, pp. 2481-2495. 10.1109/TPAMI.2016.2644615 28060704 DOI
2 
Dash, J. P., Watt, M. S., Pearse, G. D., Heaphy, M. and Dungey, H. S. (2017). "Assessing very high resolution UAV imagery for monitoring forest health during a simulated disease outbreak." ISPRS Journal of Photogrammetry and Remote Sensing, ISPRS, Vol. 131, pp. 1-14. 10.1016/j.isprsjprs.2017.07.007 DOI
3 
Derczynski, L. (2016). "Complementarity, F-score, and NLP evaluation." Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16), European Language Resources Association (ELRA), pp. 261-266.
4 
Do, M. S., Lim, E. T., Chae, J. H. and Kim, S. H. (2018). "Accuracy evaluation and terrain model creation of urban space using unmanned aerial vehicle system." The Journal of the Korea Institute of Intelligent Transport Systems, KITS, Vol. 17, No. 5, pp. 117-127 (in Korean). 10.12815/kits.2018.17.5.117 DOI
5 
Dyson, J., Mancini, A., Frontoni, E. and Zingaretti, P. (2019). "Deep learning for soil and crop segmentation from remotely sensed data." Remote Sensing, MDPI, Vol. 11, No. 16, pp. 1859. 10.3390/rs11161859 DOI
6 
Giuffrida, G., Meoni, G. and Fanucci, L. (2019). "A YOLOv2 convolutional neural network-based human-machine interface for the control of assistive robotic manipulators." Applied Sciences, MDPI, Vol. 9, No. 11, pp. 2243. 10.3390/app9112243 DOI
7 
Han, S. H. (2019). "Project design plan for drone photogrammetry." Journal of the Korean Society of Civil Engineers, KSCE, Vol. 39, No.1, pp.239-246 (in Korean).
8 
Im, S. H., Hassan, S. I., Minh, D. L., Min, K. B. and Moon, H. J. (2018). "Analysis of fusarium wilt based on normalized difference vegetation index for radish field images from unmanned aerial vehicle." The Transactions of The Korean Institute of Electrical Engineers, KIEE, Vol. 67, No. 10, pp. 1353-1357.
9 
Kim, D. I., Song, Y. S., Kim, G. H. and Kim, C. W. (2014). "A study on the application of UAV for Korean land monitoring." Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography, KSGPC, Vol. 32, No. 1, pp. 29-38 (in Korean). 10.7848/ksgpc.2014.32.1.29 DOI
10 
Kim, E. N. and Kim, D. Y. (2008). "An investigation of pine wilt damage by using ground remote sensing technique." Journal of the Korean Association of Regional Geographers, KRG, Vol. 14, No. 1, pp. 84-92 (in Korean).
11 
Kim, J. B., Jo, M. H., Oh, J. S., Lee, K. J. and Park, S. J. (2001). "Extraction method of damaged area by pine wilt disease (Bursaphelenchus xylophilus) using remotely sensed data and GIS." In Proceedings of ACRS 2001, 22nd Asian Conference on Remote Sensing, Singapore.
12 
Kim, J. B., Kim, D. Y. and Park, N. C. (2010). "Development of an aerial precision forecasting techniques for the pine wilt disease damaged area based on GIS and GPS." Journal of the Korean Association of Geographic Information Studies, KAGIS, Vol. 13, No. 1, pp. 28-34 (in Korean).
13 
Kim, J. H., Seo, I. G. and Park, S. J. (2016). "An analysis on the situation of collection and utilization of the trees damaged by pine wilt disease." Journal of Environmental Science International, KENSS, Vol. 25, No. 1, pp. 127-134 (in Korean). 10.5322/JESI.2016.25.1.127 DOI
14 
Kim, M. J., Bang, H. S. and Lee, J. W. (2017). "Use of unmanned aerial vehicle for forecasting pine wood nematode in boundary area: A case study of Sejong Metropolitan Autonomous City." Journal of Korean Society of Forest Science, KFS, Vol. 106, No. 1, pp. 100-109 (in Korean).
15 
Kim, S. R., Kim, E. S., Nam, Y. W., Choi, W. I. and Kim, C. M. (2015). "Distribution characteristics analysis of pine wilt disease using time series hyperspectral aerial imagery." Korean Journal of Remote Sensing, KSRS, Vol. 31, No. 5, pp. 385-394 (in Korean). 10.7780/kjrs.2015.31.5.3 DOI
16 
Kotsiantis, S., Kanellopoulos, D. and Pintelas, P. (2006). "Handling imbalanced datasets: A review." GESTS International Transactions on Computer Science and Engineering, Vol. 30, No. 1, pp. 25-36.
17 
Lee, H. D. (2017). Identification of damaged trees by pine wilt diseases using drone images and GIS, Master's thesis, Graduate School of Kyungil University (in Korean).
18 
Lee, J. B., Kim, E. S. and Lee, S. H. (2014). "An analysis of spectral pattern for detecting pine wilt disease using ground-based hyperspectral camera." Korean Journal of Remote Sensing, KSRS, Vol. 30, No. 5, pp. 665-675 (in Korean). 10.7780/kjrs.2014.30.5.11 DOI
19 
Lee, S. K., Park, S. J., Baek, G. M., Kim, H. B. and Lee, C. W. (2019). "Detection of damaged pine tree by the pine wilt disease using UAV Image." Korean Journal of Remote Sensing, KSRS, Vol. 35, No. 3, pp. 359-373 (in Korean).
20 
Nagai, M., Chen, T., Shibasaki, R., Kumagai, H. and Ahmed, A. (2009). "UAV-borne 3-D mapping system by multisensor integration." IEEE Transactions on Geoscience and Remote Sensing, IEEE, Vol. 47, No. 3, pp. 701-708. 10.1109/TGRS.2008.2010314 DOI
21 
Redmon, J. and Farhadi, A. (2017). "YOLO9000: better, faster, stronger." In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pp. 7263-7271. 10.1109/CVPR.2017.690 DOI
22 
Redmon, J., Divvala, S., Girshick, R. and Farhadi, A. (2016). "You only look once: Unified, real-time object detection." In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, IEEE, pp. 779-788. 10.1109/CVPR.2016.91 DOI
23 
Rokhmana, C. A. (2015). "The potential of UAV-based remote sensing for supporting precision agriculture in Indonesia." Procedia Environmental Sciences, ELSEVIER, Vol. 24, pp. 245-253. 10.1016/j.proenv.2015.03.032 DOI
24 
Smigaj, M., Gaulton, R., Suรกrez, J. C. and Barr, S. L. (2019). "Canopy temperature from an Unmanned Aerial Vehicle as an indicator of tree stress associated with red band needle blight severity." Forest Ecology and Management, ELSEVIER, Vol. 433, pp. 699-708. 10.1016/j.foreco.2018.11.032 DOI
25 
Son, M. H., Lee, W. K., Lee, S. H., Cho, H. K. and Lee, J. H. (2006). "Natural spread pattern of damaged area by pine wilt disease using geostatistical analysis." Journal of Korean Society of Forest Science, KFS, Vol. 95, No. 3, pp. 240-249 (in Korean).
26 
Xu, C. X., Lim, J. H., Jin, X. M. and Yun, H. C. (2018). "Land cover mapping and availability evaluation based on drone images with multi-spectral camera." Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography, KSGPC, Vol. 36, No. 6, pp. 589-599 (in Korean).