| الوصف: |
The prediction of rainfall remains a challenging problem, particularly with regards to seasonal forecasts that span several months. This challenge is due to the loss of atmospheric information that occurs within a few weeks, reducing the predictability of continuous seasonal forecasts. To overcome this limitation, the relationship between rainfall and other important predictors such as ocean temperature, soil moisture and snow coverage must be analyzed. However, there is difficulty to statistically analyze these elements on through observations alone, due to the limited number of events. This study proposes a new method for statistically obtaining the relationship between rainfall and predictors through the simulation results of over 40 different numerical models. A multi-layer convolutional neural network was employed to extract features from more then 6,000 years of training data. This approach is superior to classical methods such as SVD, as it enables the extraction of more nonlinear relationships among variables. The results showed that the accumulated precipitation for the rainy season(May to October) can be predicted with a 90% probability in April, with a resolution of 5°. Furthermore, the reliability of the prediction can be verified using classified loss function. Prediction of extreme drought and precipitation remains challenging due to the limited training data available, Nevertheless, by setting a threshold, forecasts can be made within that range. Further research is needed to investigate the correct method of setting the threshold value and to apply transfer learning to real data.
The 28th IUGG General Assembly (IUGG2023) (Berlin 2023) |
