Seminar Mechanics of Machines and Mechanisms - Models and Mathematical Methods

PROGRAM

Predavanja možete pratiti i online putem MITEAM stranice Seminara Mehanika mašina i mehanizama - modeli i matematičke metode: https://miteam.mi.sanu.ac.rs/asset/PgqjStRApvcGxwtBx

Plan rada Seminara Mehanika mašina i mehanizama - modeli i matematičke metode za FEBRUAR 2026.

Utorak, 24.02.2026. u 17:00, sala 301f, Kneza Mihaila 36 i Live stream
Dušan Nikezić, “VINČA” Institute of Nuclear Sciences-National Institute of the Republic of Serbia, University of Belgrade
FORECASTING AEROSOL OPTICAL THICKNESS USING DEEP LEARNING MODELS FROM NASA SATELLITE IMAGERY
Because the movement and dispersion of aerosols are related to weather patterns, it is useful to understand the movement of aerosols in order to predict weather patterns. The need for atmospheric spatio-temporal forecasts has led to the development of many well-known mathematical and physical models. Modeling and predicting approximations of complex and nonlinear functions do not provide exact solutions (ground truth). Machine Learning and Artificial neural networks with Big Data offer the new possibility to capture spatial, temporal and spatio-temporal correlations more rapidly and accurately, revealing new patterns of atmospheric non-linearity. Deep learning models for classification are easier to compare because they use the accuracy metric. Deep learning models for time series prediction is not easy to evaluate precisely, and therefore is hard to make comparison.
An overview of developed deep learning models for global spatiotemporal prediction of time series satellite imagery will be presented, followed by a discussion. There is several deep learning methods used for sequence-to-sequence prediction. For global forecasting of Aerosol Optical Thickness (AOT) from satellite imagery, the following models have been developed: ConvLSTM, CNN-LSTM, ConvLSTM-SA (self-attention mechanism), transfer learning technique with ResNet3D-101, and symmetric U-Net. The input data was MODAL2_E_AER_OD dataset of global AOT snapshots every 8 days from NASA Terra/MODIS (Moderate Resolution Imaging Spectroradiometer) satellite. Furthermore, two evaluation criteria are defined, the distance domain criterion and the time domain criterion. Since there is no exact solution in prediction, even with complicated mathematical-physical models, the obtained results showed that all the developed models are capable of finding patterns in the time domain, and can therefore be used for global AOT prediction.

Seminar Mehanika mašina i mehanizama - modeli i matematičke metode započeo je sa radom u junu 2018.god. Seminar se održava do dva puta mesečno, utorkom u periodu od 17.00 - 19.00 u Matematičkom institutu SANU.