Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Titre : Drought monitoring and forecasting through multi-sensor satellite data fusion using machine learning approaches

다중위성영상 융합을 이용한 머신러닝 기반의 가뭄 모니터링 및 예측

Auteur : Park, Seonyoung

Université de soutenance : Ulsan National Institute of Science and Technology

Grade : Doctor of Philosophy (PhD) 2018

Résumé
This thesis seek to 1) better understand the characteristics of drought factors in different climate regions (arid and humid regions), 2) monitor drought effectively using improved soil moisture to high resolution data ( 1 km) by developing drought index, 3) predict drought by identifying the relationship between drought and climate variability. In this dissertation, there are five chapters. Chapter 1 summarizes the background of research and overviews of the thesis research. In Chapter 2, the characteristics of drought factors in different climate regions were understood based on relative variable importance which is provided from machine learning approaches. Three machine learning approaches (Random forest, Boosted regression trees, and Cubist) were used to target Standardized Precipitation Index (SPI) and crop yield, and Random forest outperformed the other approaches. The variable importance targeting SPI and crop yield was applied as weight factors to produce drought maps. Those drought maps well monitored drought comparing U.S. Drought Monitor (USDM). In Chapter 3, Soil moisture with coarse spatial resolution (25 km) was downscaled to high resolution (1 km), and drought index (High resolution Soil Moisture Drought Index ; HSMDI) was developed using high resolution soil moisture. HSMDI was evaluated for three types of droughts (meteorological drought, agricultural drought, and hydrological drought). The index well monitored meteorological and agricultural drought, but there is limitation in monitoring hydrological drought. In Chapter 4, short-term (one pentad) prediction of drought conducted over East Asia by integrating remote sensing data and climate variability. The real-time multivariate (RMM) Madden–Julian oscillation (MJO) indices were used because the MJO is intraseasonal climate variability. MJO improved the performance of prediction model. Chapter 5 provides a brief summary of this study. Chapter 6 discusses future work of this study.

Présentation

Version intégrale