Deep Transfer Model Based Local Climate Zone Classification Using SAR/Multispectral Images

The deep transfer model described in this research is made up of several sub-networks that are individually optimized by an unsupervised consistency loss and a supervised task-oriented loss. Annotations are used by the supervised loss function to achieve the intended outcome. The consistency loss promotes various sub-networks to communicate gained knowledge and motivates the network to learn the target domain data distribution. We employ the suggested model to deal with a classification of local climate zones worldwide. The dataset comprises 48,307 samples from 10 additional cities in the target domain and 352,366 training samples from 42 locations in the source domain. We provide a unique method for local climate zone (LCZ) categorization utilizing deep learning and big data analytics to merge publically available global radar and multi-spectral satellite data, gathered by the Sentinel-1 and Sentinel-2 satellites. A consistent classification scheme for characterizing the local physical structure worldwide is provided by LCZ. Due to the rapid advancement of satellite imaging techniques, multispectral (MS) and synthetic aperture radar (SAR) data have become increasingly common in LCZ classification jobs. As demonstrated by our experiments; the suggested model improves the performance as compared to baseline model.