|
Comparison of Machine Learning Methods for Statistical Downscaling of Climate Models Precipitation and Temperature
|
Nima Nemati Shishehgaran *   , Ali Sadian |
| Tarbiat Modares University |
|
|
Abstract: (18 Views) |
| Statistical downscaling of General Circulation Model (GCM) outputs to obtain accurate climate estimates at local scales is one of the fundamental challenges in climate change studies, particularly in hydrologically sensitive regions. Considering the limitations of conventional statistical methods and the emergence of modern machine learning algorithms, the present study aims to evaluate the performance of three methods—Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Random Forest (RF)—for downscaling precipitation and temperature variables. Accordingly, the outputs of 10 CMIP6 climate models under four scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0 and SSP5.58) were extracted for the upstream basin of the Karkheh Dam during the historical period 1990–2014. The data were used in a 20-year period for training and a five-year period for testing the models. Evaluation results based on three statistical indices (R², MAE, and RMSE) showed that the RF method outperformed the other two methods for both precipitation and temperature, while SVM yielded the weakest results, particularly in precipitation downscaling. Leveraging the superior RF approach, climate variables were projected for the periods 2031-2055 and 2056-2080, with trends analyzed using the Mann-Kendall test and Sen's slope estimator. The findings indicated that mean precipitation decreases under all four scenarios, with reductions of approximately 2.96% to 19.22%, while average temperature increases by 1.13°C to 3.13°C. Furthermore, trend analysis of precipitation and temperature variables for future periods under the evaluated scenarios indicates the persistence of natural and random fluctuations in precipitation alongside statistically significant upward trends in regional mean temperature. These results highlight the high sensitivity of the study area to climate change impacts and the necessity of employing machine-learning-based approaches to improve the accuracy of projections. Moreover, the application of these models enhances the ability to capture complex climate patterns and can effectively contribute to improving the downscaling of climate-model-simulated variables. |
Article number: 6 |
| Keywords: Climate Change, General Circulation Models, Multilayer Perceptron, Support Vector Machine, Random Forest, Upstream Karkheh River Basin |
|
|
|
|
Type of Study: Research |
Subject:
Special
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
