Multimodal learning has become increasingly relevant in environmental modelling, where heterogeneous sources such as satellite observations and meteorological measurements are jointly used to estimate urban air pollutant concentrations. Despite encouraging predictive performance, it remains unclear which modalities and temporal components truly contribute useful information, and which instead introduce redundancy or noise. Understanding these relationships is essential for designing efficient and interpretable models rather than progressively more complex architectures. This thesis proposes a multimodal framework for urban air quality forecasting over Milan. The approach integrates multi-sensor satellite imagery (Sentinel-1, Sentinel-2, Sentinel-3, and Sentinel-5), land-cover data, and a digital elevation model together with meteorological observations.