Supervised Machine Learning in Drought Evaluation

Abstract

Among natural disasters, drought is one of the most common threats to many regions of the world and to Georgia as well. The monitoring and prediction methods of drought and precipitation distribution, the possibilities of their application in the reality of Georgia are considered in proposed work. Simulation methods such as Machine Learning (ML), namely Supervised Machine Learning (SML), optimal for similar complex tasks are presented as the alternative research methods. To conduct research, 1960-2022 period data were taken from database of National Environment Agency and the reanalysis data of the 1960-1990 Copernicus ERA5 rainfall, which were compared with the data of the stations on the territory of Georgia for the validation purpose. Standardized Precipitation Index (SPI) was selected as the research parameter. Using the prediction model and algorithm, drought-vulnerable areas in the Kakheti region were identified. As a result of comparison, lowest correlation rate was 0.309 at Shiraki, maximum was 0.657 at Omalo; minimum mean absolute error 1,662 at Udabno, the maximum 3,041 in Shilda. The smallest standard deviation 4,047 was fixes at Udabno, largest 7,624 at Lagodekhi. By analyzing stations data and satellite sources, it was determined that using the regression method of Machine Learning, it is sufficient to evaluate 1960-2000 period data for learning and 2001-2022 period data for training. The training time of Bagged Trees Optimized Algorithm was recorded as 326.21 sec, prediction speed ~ 7900obs/sec, RMSE - 0.5046, R2-0.64, MSE-0.25466, MAE-0.38065, training process minimum leaf size 19, and 40 iterations are assigned for optimization. CHIRPS satellite data were taken for next generation of the model. For prediction, it was necessary to calculate linear regression equation for each station. In the first case of forecast scenario, the amount of precipitation was determined from 0 cm to 10 cm. Gurjaani was highlighted, where forecast assumed SPI value from -0.008 to -0.901, and Kvareli- the SPI value from -0.002 to -0.138.  The use of the presented ML model and algorithm for the analysis of precipitation distribution, drought monitoring and prediction is appropriate for Kakheti and other regions too in conditions of proper observation data base (60 years). It is recommended to use obtained results in early warning system for drought monitoring.