Rahman, Chowdhury MofizurKarim, Dewan ZiaulTamanna, Thufa AnwarJoy, Shah Newaz KhanMasum, ZayedAbdullah, Fariha BinteAnam, S M Kafi2026-08-092026-08-0920252026-02-06ID 21241080ID 21201591ID 22101461ID 23241050ID 24241280https://hdl.handle.net/10361/28831This thesis is submitted in partial fulfillment of the requirements for the degree of Bachelor of Science in Computer Science and Engineering, 2026.Cataloged from PDF version of thesis.Includes bibliographical references (pages 54-56).Air pollution and changes in the weather make it hard to keep an eye on the environment and protect public health, especially in areas where there aren’t many ground-based sensors or they are spread out unevenly. Ground measurements give accurate local readings, but they fail to encompass much of ground data. Spatial coverage limits analysis on a regional scale. Satellite observations provide extensive spatial data but lack local accuracy. This thesis suggests a complete multimodal machine learning framework that combines satellite-derived spatial features, ground-based environmental measurements, regional clustering, and time-series forecasting to better predict and understand air quality and important weather variables. The suggested framework uses gradient-boosted regression models for multimodal learning and pre-trained convolutional neural networks (ResNet-18, ResNet-50, and DenseNet- 121) to get features from satellite images. Time-series forecasting is used to find patterns over time and changes in environmental factors that happen at different times of the year. Unsupervised regional clustering is used to find groups of districts that have similar pollution and weather patterns. The framework is assessed against various air pollutants (SO, NO2, O3, PM2.5, PM10) and meteorological factors (solar radiation, relative humidity, and rainfall) utilizing standard performance metrics, such as RMSE, MAE, and R2. The experimental results demonstrate the effectiveness of multi-modal fusion is highly dependent on the target and region. For variables that were affected, like sulfur dioxide and rainfall, explained variance went up a lot and prediction error went down a lot. On the other hand, O3 and solar radiation got a little better. Again, NO2, particulate matter, and relative humidity were best modeled using only ground-based data. This suggests that there are many localized emission sources and processes that happened close to the ground. To improve the transparency of the model, explainable artificial intelligence (XAI) methods were added through global feature importance analysis with XGBoost. This gave us a better idea of what environmental factors were most important for making predictions over time. The results underscore the significance of pollutant- and region-specific modeling approaches and illustrate that multimodal learning can enhance environmental prediction when informed by the physical and statistical attributes of the target variables.60 pagesen-USAttribution-NonCommercial-NoDerivatives 4.0 InternationalBRAC University theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission.http://creativecommons.org/licenses/by-nc-nd/4.0/Machine learningClimate changeEnvironmental monitoringSatellite imageryTime series forecastingData mappingExplainable AIArtificial intelligence--Environmental applications.Time-series analysis.Neural networks (Computer science).Environmental monitoring--Remote sensing.Geospatial data.Thesis