연구보고서
- 저자
- 이진영 박사
- 작성일
- 2016.01.23
- 조회
- 251
- 요약
- 목차
Drought is a regularly occurring natural disaster that affects large numbers of people and causes serious economic damage. The frequency of severe drought is expected to increase in many regions due to the effect of climate change. Despite the high levels of uncertainty on the spatial extent and the temporal timing of the changes monitoring the current conditions and forecasting future drought several months ahead can be valuable to reduce the negative impacts. It is essential to develop a drought early warning system that performs drought monitoring and forecasting, delivers the information to decision-makers in a timely manner, and results in the reduction of the adverse impacts of drought.
Many hydro-meteorological variables should be included for comprehensive drought monitoring and forecasting. Since observation data of the variables are only available where there are dense networks of weather stations, remotely sensed data can be used instead. The purpose of this study is to develop drought monitoring and forecasting techniques using remotely sensed data to be used for regions where observation data is lacking or absent. In this study, hydrometeorological and biophysical variables such as precipitation, land surface temperature, actual and potential evapotranspiration, fractional photosynthetically active radiation (FPAR), leaf area index (LAI), vegetation index, and soil moisture were obtained from remotely sensed data. Several variables of actual and potential evapotranspiration and soil moisture were also derived from land surface model outputs. The large-scale climate indices Arctic Oscillation (AO) and Multivariate ENSO Index (MEI) were integrated with these variables to obtain drought information.
While drought monitoring can be traditionally performed based on drought indices such as Standardized Precipitation Index (SPI) and Palmer Drought Severity Index, a machine learning technique, random forest, was used in this report in developing drought monitoring and forecasting models to overcome the limitation of short records. Drought-related variables are intertwined though complex feedbacks and couplings between variables, and the interactions between them present challenges in simulations using simple algorithms. Including limited number of variables may lead to inaccurate results of drought conditions. The variables derived from remotely sensed data may be less accurate than observation data since in many cases they are estimated from many algorithms which are physically based but far simpler than real-world interactions. In this case, machine learning techniques can be effectively used to develop drought monitoring and forecasting models since they are not required to explicitly provide or to make assumptions on physical relationships between variables.
In this report, random forest was used to develop drought monitoring and forecasting models with a lead time from 1-6 months. The random forest model is known to produce robust results against outliers and noises. As input variables, variables derived from remotely sensed data (day, night, and mean soil moisture; day, night and mean land surface temperature; actual and potential evapotranspiration; vegetation indices of Normalized Difference Vegetation Index and Enhanced Vegetation Index; precipitation; FPAR; LAI), variables derived from land surface model output (actual and potential evapotranspiration; soil moisture), and large-scale climate indices (AO and MEI) were used. As target variables of the models, SPI and Standardized Precipitation Evapotranspiration Index based on observation data as well as Normalized Difference Water Index (NDWI), Normalized Difference Drought Index (NDDI), and Normalized Multi-band Drought Index (NMDI) based on remotely sensed data were used.
Remotely sensed and observation data were compared for precipitation, land surface temperature, potential evapotranspiration, and soil moisture. For precipitation, TRMM monthly rainfall data produced relatively small errors and high correlations in the regions with lower latitudes. There also exist high correlations between MODIS land surface temperature and observed air temperature except for daytime land surface temperature during summer. While the MODIS potential evapotranspiration appeared closer to the calculated values based on Penman-Monteith compared to the land surface model output, the soil moisture data from the land surface model output were closer to the KoFlux observation data than the remotely sensed data from AMSR-E sensor. A regression model was built between the AMSR-E derived data and the land surface output for soil moisture, and it was applied to AMSR-E derived soil moisture data to enhance the spatio- temporal resolutions of the remote sensing-based soil moisture data and to reduce estimation errors.
Drought monitoring models with several sets of input variables were developed using random forest, and their performances were evaluated. A drought monitoring model with all available variables performed the best. Random forest provides internal estimates of generalization errors and relative variable importance using out-of-bag procedures. Precipitation turned out to be exceptionally important for target variables based on observation data. For target variables based on remotely sensed data, the relative importance of large-scale climate indices of AO and MEI were very high. It indicates that drought conditions in the study area are very much affected by teleconnections. The correlations between input data and output of models for NDWI6 (NDWI with MODIS band 6) and NDWI7 (NDWI with MODIS band 7) were especially high, suggesting them as useful target variables for drought monitoring.
Drought forecasting models with a lead time from 1-6 months were developed using input variables previously selected for drought monitoring models, and their performances were evaluated. Forecasting errors tend to increase with longer lead times, while the errors tend to decrease with more past months of the target variable used as input data. The number of past months of the target variable affects forecasting errors more than the lead time used. Target variables NDDI5 (NDDI with MODIS band 5) and NDDI6 (NDDI with MODIS band 6) were excluded because of their unstable error levels. Drought forecasting models developed performed well for other target variables. The target variable itself with past values showed the highest relative importance among input variables. Large-scale climate indices of AO and MEI also have large relative importance as with drought monitoring models. The relative importance of vegetation-related variables of FPAR, LAI, NDVI, and EVI were moderately low.
The drought forecasting models developed using target variables based on remote sensing were applied for the crop yield data of highland bok choy and radish as well as frequency and area data of forest fire. Rather than developing models for directly forecasting crop yields or forest fire frequencies, the statistical correlations between crop yields or forest fire frequencies and remote sensing-based target variables for administrative districts of the study area were obtained, and then the performance of the models for the administrative districts with high correlations was evaluated on the remote sensing-based target variables.
Among nine SI-DO and six SI-KUN-KU used for the crop yield analyses, crop yield data and remote sensing-based target variables showed high correlations in two SI-DO of Jeollabuk-do and Gyeongsangnam-do and four SI-KUN-KU of Yeongwol-gun, Yanggu-gun, Inje-gun, and Gangneung-si for highland bok choy. High correlations were observed for highland radish in three SI-DO of Gangwon-do, Gyeongsangbuk-do, and Gyeongsangnam-do, as well as five SI-KUN-KU of Taebaek-si, Samcheok-si, Yeongwol-gun, Inje- gun, and Gangneung-si. Relatively low correlations were observed in some cases since there are many other factors affecting crop yield. The remote sensing-based target variables, however, can be used as good proxies for the cases previously listed.
The performance of a drought forecasting model using NDWI7 as a target variable with 6-month lead time was especially good in Yeongwol- gun for highland bok choy, as well as a model using NDWI7 with 1-month lead time in Gyeongsangnam-do, and a model using NDWI6 with 2-month and 6-month lead times respectively in Yeongwol-gun. For highland radish, models using NDWI6 with 6-month lead time in Gyeongsangnam-do and 2-month lead time in Gyeongsangbuk-do showed high correlations with crop yield.
The correlations between the remote sensing-based target variables and the forest fire frequency data, fire area data, and the fire area ratio to the total forest area were tested and only forest fire frequency showed good correlations. The high correlations were only observed when analyzed for SI-DO level, and for NDWI and NDDI among many target variables. The correlations were especially high in Busan, Ulsan, Chungcheongbuk-do, Jeollanam-do, Gyeongsangbuk-do, and Gyeongsangnam-do, and the performance of the drought forecasting models for those administrative districts for forecasting the target variables were excellent in all cases.
The developed drought monitoring and forecasting models based on a machine learning methodology, random forest, can be used by a variety of end users. The models developed with target variables based on observation data with time scales from 1-9 months can be applied to many different types of drought. The models are suited to provide general drought information for many drought-affected sectors by being integrated in the drought information system operated by Korea Meteorological Administration. Since the NDWI and NDDI were developed targeting drought impacts on vegetation, and the NMDI was developed for monitoring soil and vegetation water conditions, the models developed with target variables of remote sensing-based drought indices of NDWI, NDDI, and NMDI can provide the most valuable information for agricultural drought (soil moisture drought). As applied with highland crop yield data or forest fire frequency data, the models can provide valuable information for farmers and forest managers by being integrated with drought information system operated by the Ministry of Agriculture, Food and Rural Affairs or the Korea Forest Service.
Since remote sensing has not yet been actively used for drought monitoring and forecasting, the integration of remote sensing techniques to existing drought information system or early warning systems operated by many government organizations offers a large potential. An environment where drought information derived from remote sensing can be used by end- users with confidence should be built soon, by performing a range of case studies for evaluating remote sensing data and for improving their reliability.