Self-Ensemble Semi Supervised Learning Based SAR and Multispectral Local Climate Zone Classification

Local climate zones are essential for understanding urban climates and their effect on environmental processes. Here we demonstrate self-ensemble learning based approach to classify local climate zone. The big So2Sat LCZ42 benchmark dataset has been used in this paper. The dataset contains SAR and multispectral data from sentinel 1 and sentinel-2 satellites. SAR has the capability to penetrate though cloud cover and provide surface structure and texture information, whereas multispectral data provides spectral information to distinguish waterbodies, vegetation and built-up areas. Semi supervised learning using Student-Teacher network and Student-Student has been implemented in this paper to increase the accuracy achieved by baseline neural network. Student-Teacher network (ST-Net) use exponential moving average, and that's why the teacher network is tightly coupled with student network. we use Student-Student network (SS-Net) by replacing teacher with an independent student. The SS-Net is not tightly coupled and hence performs better than baseline (Resnet) and ST-Net.