Geolocation assisted street view imagery self-supervised learning for urban land use analysis

Self-supervised learning (SSL) can fully utilize the massive amount of unlabelled samples in geoscience and remote sensing fields, thereby significantly reducing the data annotation cost required for large models. Unlike expensive manually annotated semantic labels, geolocation information of these unlabelled samples is usually available for free; however, it has not been effectively utilized in conventional SSL methods. This study aims to fully exploit the geolocation information inherent in street view images that are freely accessible through mainstream map platforms to enhance self-supervised pre-training of image classification models, thereby improving performance on the downstream task of urban land use analysis. Specifically, in this letter, this geolocation information is used together with image features obtained from SSL to screen unlabelled samples and generate pseudo-labels for them. Leveraging these high-quality pseudo-labelled samples, a novel two-stage progressive fine-tuning strategy is proposed to optimize the feature representation of SSL. The method is independent of specific model architectures and SSL approaches, demonstrating generalizability. Experimental results show that the proposed method achieves state-of-the-art performance in street view imagery-based urban land use analysis tasks such as urban functional zone identification and urban ground object classification among SSL-based methods and outperforms the mainstream supervised transfer learning approach.