| Abstract [eng] |
The aim of this Master's thesis is to investigate the applicability of deep machine learning methods to short-term precipitation nowcasting using radar data. During the study, 15 different deep machine learning models were implemented, including convolutional and recurrent neural networks and their variations. Different data subsets were used to train and predict the models from a dataset of radar precipitation images of the Lithuanian region, which covers the period from 1 March 2022 to 3 December 2023. The results show that deep machine learning models can be efficiently applied to short-term precipitation nowcasting. The best forecasting performance was achieved by the UNet and ConvLSTM architectures, with F1-accuracy estimates above 0.85 and root mean square errors below 0.04. The study also showed that the application of convolutional neural network localisation techniques can significantly improve the performance of the models in forecasting localised meteorological events. This is particularly important for the analysis of regional precipitation trends. The results section of the study provides a comparative analysis of the developed models, detailing their capabilities and limitations in forecasting various meteorological phenomena, and assesses the usefulness of localisation methods in the field of analysis. |
