Meteorological drought forecasting using decision tree

Drought is a natural phenomenon that directly affects human life in a region. Therefore, drought classification and prediction models are essential tools for the mitigation of adverse consequences of drought. Beginning with a brief review of different machine learning techniques to predict drought, the applicability of decision tree for drought classification and the prediction was investigated in this thesis. For this aim, quantitative values of Standardized Precipitation Evapotranspiration Index (SPEI) were collected from a global SPEI data repository (grid points across Antalya province) during the period of 1961-2015. A dataset including 659 samples was created and considered as features in this study. The target SPEI values were first calculated by averaging SPEI values of nearby grid points. Then, drought classes for target SPEI values were determined in terms of wet and dry events. It is well documented that negative monthly SPEI values in a region generally represent dry months at the given region. The lower the SPEI, severe the drought. A binary classification decision tree model was presented that uses the SPEI values of each grid point at month t as inputs and predicts the drought classes (wet or dry) of Antalya at one month later (i.e., month t+1). The dataset was divided into train and validation sets. The predicted and observed values of drought classes were compared using performance evaluation criteria such as total accuracy, kappa statistics, and classification error. It was obtained that a value of 80.63, 19.37, and 0.61 for total accuracy, classification error, and kappa statistics, respectively. This indicated that the presented decision tree model is moderately accurate for drought prediction in the case study region. On the other hand, results showed an increasing trend for dry events in the most recent years that can be considered as the signal for climate variability in the study area.