SAR and Multispectral Image Fusion using Multibranch CNN and Cross Domain Learning for Local Climate Zone Classification

We provide a novel method for local climate zone (LCZ) classification utilizing deep learning and big data to merge publicly available SAR and multispectral (MS) data, gathered by the Sentinel-1 and Sentinel-2 satellites. LCZ provides a unified classification scheme to explain the physical layout of objects on a worldwide scale. Due to the rapid advancements in satellite imaging techniques, MS and SAR data have become increasingly popular for use in LCZ classification. The problems of diverse features and different imaging techniques continue to arise when combining MS and SAR data. Our method is based on fusion of SAR and MS data using multi branch CNN and novel cross domain feature learning. The experiments were performed using So2Sat LCZ42 dataset, and the results show that our model outperforms SAR only and MS only classification. During the training phase, the model parameters are modified to enhance the interpretability of patterns pertaining to climate in urban regions. A validation dataset and evaluation criteria such as output accuracy, kappa accuracy, average accuracy is used to analyze the model. The findings are useful for legislators, urban planners, environmental scientists, and other professionals involved in climate-resilient urban planning. The findings demonstrate the use of multispectral data fused with SAR data for precise LCZ categorization, contributing to our understanding of the unpredictable nature of urban climates and assisting planners in making well-informed decisions regarding urban development strategies.