Deep learning techniques have garnered significant attention in remote sensing scene classification. However, obtaining a large volume of labeled data for supervised learning (SL) remains challenging. Additionally, SL methods frequently struggle with limited generalization ability. To address these limitations, self-supervised multi-mode representation learning (SSMMRL) is introduced for local climate zone classification (LCZC). Unlike conventional supervised learning methods, SSMMRL utilizes a novel encoder architecture that exclusively processes augmented positive samples (PSs), eliminating the need for negative samples. An attention-guided fusion mechanism is integrated, using positive samples as a form of regularization. The novel encoder captures informative representations from the unannotated So2Sat-LCZ42 dataset, which are then leveraged to enhance performance in a challenging few-shot classification task with limited labeled samples. Co-registered Synthetic Aperture Radar (SAR) and Multispectral (MS) images are used for evaluation and training. This approach enables the model to exploit extensive unlabeled data, enhancing performance on downstream tasks. Experimental evaluations on the So2Sat-LCZ42 benchmark dataset show the efficacy of the SSMMRL method. Our method for LCZC outperforms state-of-the-art (SOTA) approaches.

Deep learning has demonstrated significant potential in remote sensing scene classification; nevertheless, its efficacy is frequently hindered by the requirement for extensive labeled datasets and its restricted generalisation in practical applications. We developed a Self Supervised multimodal representation Learning (MMRL) framework for Local Climate Zone Classification (LCZC) to tackle these problems. Our approach utilises novel encoder architecture that derives representations exclusively from synthetic aperture radar (SAR) and multispectral (MS) data. To enhance the learning process, we implemented multimode loss and consistency loss, enhancing the model ability to prioritize the most meaningful features. The encoder, trained on the unlabelled So2Sat-LCZ42 dataset, acquires robust and transferable representations, further refined for a few-shot LCZ classification task utilising a restricted number of labelled samples. This approach allows the model to efficiently utilize extensive unlabeled data, leading to enhanced performance in subsequent classification tasks. Experimental findings on the So2Sat-LCZ42 benchmark validate that our self supervised learning (SSL) based approach achieved better accuracy than the current state-of-the-art(SOTA) SSL models, illustrating its efficacy for label-efficient LCZC.