Fusion and Grouping Strategies in Deep Learning for Local Climate Zone Classification of Multimodal Remote Sensing Data

Researchers Ancymol Thomas and Jaya Sreevalsan-Nair have published a comprehensive analysis of fusion strategies for deep learning-based Local Climate Zone (LCZ) classification, addressing a critical gap in how multimodal remote sensing data is combined for urban climate analysis. Their study, conducted exclusively on the So2Sat LCZ42 dataset containing Synthetic Aperture Radar (SAR) and Multispectral Imaging (MSI) pairs, systematically tested four different fusion models: a baseline hybrid fusion (FM1), attention-enhanced fusion (FM2), multi-scale Gaussian filtered fusion (FM3), and weighted decision-level fusion (FM4). The research represents a significant methodological advancement in urban climate informatics, providing a framework for more accurate zoning maps that study urban structures and land use impacts.

The team's ablation experiments revealed that their baseline hybrid fusion model (FM1) consistently outperformed more complex approaches, particularly when combined with band grouping within data modalities and label merging in ground truth. This optimized configuration achieved 76.6% overall accuracy, demonstrating superior performance for underrepresented classes that often challenge traditional classification methods. The study's practical implications are substantial—urban planners and climate scientists can now leverage these fusion strategies to create more accurate LCZ maps from freely available satellite data, enabling better analysis of urbanization's impact on local climate patterns and supporting evidence-based urban planning decisions worldwide.