연구보고서
- 저자
- 이진영 박사
- 작성일
- 2017.07.04
- 조회
- 205
- 요약
- 목차
Three regions of Indonesia with different rainfall characteristics were chosen to develop drought forecast models based on machine learning. The 6-month Standardized Precipitation Index was selected as the target variable. The models’ forecast skills were compared to the skills of long-range climate forecast models in terms of drought accuracy and regression-based Mean Absolute Error (MAE).
Indonesian droughts are known to be related to El Nino Southern Oscillation (ENSO) variability despite regional differences, as well as related to monsoon, local Sea Surface Temperature (SST), other large-scale atmosphere-ocean interactions (such as the Indian Ocean Dipole and Southern Pacific Convergence Zone), and local factors including topography and elevation. Machine learning models were thus utilized to enhance drought forecast skill by combining local and remote SST and remote sensing information, reflecting intial drought conditions to the long-range climate forecast model results.
A total of 126 machine learning models were developed and six long-range climate forecast models (MSC_CanCM3, MSC_CanCM4, NCEP, NASA, PNU, POAMA) with 1 to 6-month lead times, as well as one climatology model based on remote sensing precipitation data were used for the three regions of West Java, West Sumatra, and Gorontalo.
When comparing the results of the machine learning models and long-range climate forecast models, the West Java and Gorontalo regions showed similar characteristics in terms of drought accuracy. Drought accuracy of the long-range climate forecast models were generally higher than the machine learning models with short lead times, but the opposite occurred for longer lead times. For West
Sumatra, however, the machine learning models and the long-range climate forecast models showed similar drought accuracy. The machine learning models showed smaller regression errors for all three regions, especially with longer lead times.
Among the three regions, the machine learning models developed for the Gorontalo region showed the highest drought accuracy and the lowest regression error. The West Java region showed higher drought accuracy compared to the West Sumatra region, while West Sumatra showed lower regression error compared to West Java. The lower error in the West Sumatra region may be due to the smaller sample size used for training and evaluation for the region. Regional differences in forecast skill are determined by the effect of ENSO and the following forecast skill of the long-range climate forecast models.
Relative importance of input variables of the machine learning models were evaluated for each region. In all cases, the importance of the SPI6 forecast (SPI6 FCST) was the highest, especially with short lead times. For West Java, month (MONTH) and the SST forecast of the Nino3 region (NINO3_SST FCST) showed higher importance. For West Sumatra, locational information including latitude (LAT) and longitude (LON) has higher importance, probably due to the shape of the region. For Gorontalo, the importance of SST forecast in SPCZ (SPCZ_SST FCST) was more obvious compared to West Java. While somewhat high in West Sumatra, the relative importance of remote sensing variables were generally low in most cases. High importance of the variables based on long-range climate forecast models indicates that the forecast skill of the machine learning models are mostly determined by the forecast skill of the climate models.
A case study was performed for West Java. Drought forecast results for drought events during 2003-2015 based on the NASA model, indicate that the long-range forecast models tend to underestimate drought severity, while the machine learning models tend to forecast drought to a narrower range of drought index values. It may be solved by testing the design of machine learning model training.
If given information on drought occurrences in advance, loss and damages due to drought can be reduced by securing resources and planning its efficient allocation. In order to prevent and reduce drought risk, nonstructural measures such as the development of a drought monitoring and forecasting system is essential. Through this study, forecast skill of meteorological drought in Indonesia can be improved. The drought forecast models developed in this study can help to reduce drought disaster risk.