°ÇÃ൵½Ã°ø°£¿¬±¸¼Ò

Architecture & Urban Research Institute

pdf¿ø¹®º¸±â ¿¡·¯ ÇØ°á¹æ¹ý ¹Ù·Î°¡±â



¹®ÇåȨ > Àú³Î > »ó¼¼

´ëÇÑÁö¸®ÇÐȸ|´ëÇÑÁö¸®ÇÐȸÁö 2018³â 6¿ù

±â»ç¸í ¼­¿ï ¹Ì¼¼¸ÕÁö µ¥ÀÌÅÍ °áÃø´ëÄ¡¸¦ À§ÇÑ ½Ã°ø°£ Å©¸®±ëÀÇ ¾Ó»óºí ¸Ó½Å·¯´× / An Ensemble Machine Learning from Spatio-temporal Kriging for Imputation of PM10 in Seoul, Korea
ÀúÀÚ¸í ¼ÛÀλó ; ÀÌâ·Î ; ¹Ú±âÈ£
¹ßÇà»ç ´ëÇÑÁö¸®ÇÐȸ
¼ö·Ï»çÇ× ´ëÇÑÁö¸®ÇÐȸÁö , Vol.53 No.3(2018-06)
ÆäÀÌÁö ½ÃÀÛÆäÀÌÁö(427) ÃÑÆäÀÌÁö(18)
ISSN 1225-6633
ÁÖÁ¦ºÐ·ù µµ½Ã
ÁÖÁ¦¾î °áÃø´ëÄ¡ ; ½Ã°ø°£ Å©¸®±ë ; ÀçÇ¥Áý ; ¹Ì¼¼¸ÕÁö ; Imputation ; Spatio-temporal kriging ; Resampling ; Particulate matter
¿ä¾à1 ½Ã°ø°£ µ¥ÀÌÅÍÀÇ °áÃøÄ¡´Â ±× ÀÚü·Î µ¥ÀÌÅÍÀÇ °áÇÔÀ¸·Î¼­ ½Ã°ø°£ ºÐ¼® °á°ú¸¦ ¿Ö°î½Ãų ¼ö ÀÖ´Ù. ±×·¯³ª ½Ã°ø°£ µ¥ÀÌÅÍ¿¡ ³»ÀçµÈ ½Ã°ø°£ ÀÇÁ¸¼ºÀ» ÀÌ¿ëÇÑ °áÃø´ëÄ¡ ¹æ¹ýÀº ´ú ÁÖ¸ñ¹Þ¾Æ ¿Ô´Ù. ÀÌ¿¡ º» ¿¬±¸¿¡¼­´Â ¼­¿ïƯº°½Ã ¹× ±Ù¹æÀÇ 54°³ ÃøÁ¤¼Ò·ÎºÎÅÍ 2010³âºÎÅÍ 2014³â±îÁö 5³â°£ ÃøÁ¤µÈ ½Ã°£º° ¹Ì¼¼¸ÕÁö(PM10) µ¥ÀÌÅÍÀÇ °áÃøÄ¡¸¦ ´ëÄ¡Çϱâ À§ÇÏ¿© ¾Ó»óºí ½Ã°ø°£ Å©¸®±ë ¸ðÇü¿¡ ±âÃÊÇÑ °áÃø´ëÄ¡ ¸ðÇüÀ» Á¦¾ÈÇÏ¿´´Ù. ±âÁ¸ ¿¬±¸µéÀ» °ËÅäÇÑ °á°ú, º» ¿¬±¸¿¡¼­ ÀÌ¿ëµÈ Á¢±Ù¹ýÀÇ Çʿ伺ÀÌ ¹ß°ßµÇ¾ú´Ù. º» ¿¬±¸°¡ Á¦¾ÈÇÏ´Â ¾Ó»óºí °áÃø´ëÄ¡ ¸ðÇüÀº ´Ü±â°£ÀÇ ½Ã°ø°£ µ¥ÀÌÅÍ¿¡¼­ ÀçÇ¥Áý(resampling)µÈ ÇÏÀ§ µ¥ÀÌÅͼÂÀ¸·Î º¹¼öÀÇ ½Ã°ø°£ Å©¸®±ë ¸ðÇüµéÀ» ÀûÇÕÇÏ°í, À̵éÀ» ¾Ó»óºíÇÏ¿© °áÃø´ëÄ¡ Á¤È®µµ¸¦ ³ôÀÌ°íÀÚ ÇÑ´Ù. Çâ»ó ¿©ºÎ¸¦ ½ÇÁõÇϱâ À§ÇÏ¿© ÃøÁ¤ µ¥ÀÌÅÍ¿¡ ´ëÇØ °áÃø´ëÄ¡ ½ÇÇèÀ» ½Ç½ÃÇÏ¿´´Ù. ½ÇÇè¿¡¼­´Â ÀçÇ¥Áý Ƚ¼ö, ½Ã°ø°£ Å©¸®±ë ÀûÇÕ ½Ã ÀÌ¿ô ºñÀ², °áÃø »ý¼º ºñÀ² µî3 ¿ä¼Ò¿¡ ´ëÇØ ¼­·Î ´Ù¸¥ Á¶°ÇµéÀ» Àû¿ëÇÏ¿´´Ù. ½ÇÇè °á°ú, Á¦¾ÈµÈ ¾Ó»óºí ¸ðÇüÀº ´ÜÀÏ ½ÃÇà ½Ã°ø°£ Å©¸®±ë ¸ðÇü(1.32~11.36%)°ú, ¼±Çü ¾Ó»óºí ¸ðÇü(Æò±Õ 52%)º¸´Ù ³ôÀº Á¤È®µµ·Î °áÃøÄ¡¸¦ ´ëÄ¡ÇÏ¿´´Ù. º» °á°ú´Â Á¦ÇÑµÈ È¯°æ¿¡¼­ ½Ã°ø°£ Å©¸®±ë ¸ðÇü ¾Ó»óºíÀÌ °áÃø ´ëÄ¡ Á¤È®µµ¸¦ ³ôÀÌ´Â µ¥ È¿°ú°¡ ÀÖÀ½À» ÀÔÁõÇÑ´Ù. ´Ù¸¸ Á¦¾ÈµÈ ¾Ë°í¸®ÁòÀÇ Á¤È®¼ºÀº ¸Ó½Å·¯´× ±â¹ÝÀÇ °áÃø´ëÄ¡ ¾Ë°í¸®Áò¿¡ ºñÇؼ­ ´ú ¿ì¼öÇߴµ¥, ÀÌ °á°ú´Â ¸Ó½Å·¯´× ¾Ë°í¸®Áò¿¡¼­ ½Ã°ø°£ ÀÇÁ¸¼º È¿°ú°¡ ¾î¶»°Ô ³ªÅ¸³ª´ÂÁö¿¡ ´ëÇÑ Ãß°¡ ¿¬±¸ Çʿ伺À» Á¦±âÇÑ´Ù.
¿ä¾à2 Missing values in spatio-temporal data presumably cause defects, such that contaminate the results of spatio-temporal analyses. However, imputation methods for spatio-temporal data considering the inherent nature of spatio-temporal dependence have been neglected. We suggest an imputation algorithm based on ensemble spatio-temporal kriging for particulate matter measurement data for the period 2010-2014 at 54 monitoring stations near the metropolitan city of Seoul, Korea. We review previous studies on imputation methods for spatio-temporal data, then shed light on the necessity of our approach. Our approach implements resampling techniques on limited spatio-temporal data for a short-term period, then aims to enhance the imputation accuracy by taking the ensemble of the imputation results of resampled sub datasets. To examine such enhancement, we apply different conditions in experiments, including the number of resampling, neighborhood ratios, and ratios of artificially generated missing values. Results show that our approach outperforms both spatio-temporal kriging with the whole dataset (1.32~11.36%) and the linear regression-based imputation algorithm (52% in average). Our results show that the learning approach by resampling is still effective in spatiotemporal kriging in a limited environment as well as the spatio-temporal algorithm considering the inherent dependence among the data. But the considerable underperformance compared to the accuracy of the machine learning-based algorithm indicates the necessity of further examination of the effect of spatio-temporal dependence in such an algorithm.
¼ÒÀåó ´ëÇÑÁö¸®ÇÐȸ
¾ð¾î Çѱ¹¾î