Cross-scale Urban Land Cover Mapping: Empowering Classification through Transfer Learning and Deep Learning Integration

Cross-scale Urban Land Cover Mapping: Empowering Classification through Transfer Learning and Deep Learning Integration

Urban land cover mapping is essential for effective urban planning and resource management. Thanks to its ability to extract intricate features from urban datasets, deep learning has emerged as a powerful technique for urban classification. The U-net architecture has achieved state-of-the-art land cover classification performance, highlighting its potential for mapping urban trees at different spatial scales. However, deep learning approaches often require large, labeled datasets, which are challenging to acquire for specific urban contexts. Transfer learning addresses this limitation by leveraging pre-trained deep learning models on extensive datasets and adapting them to smaller urban datasets with limited labeled samples. Transfer learning can enhance classification performance and generalization ability. In this study, we proposed a novel cross-scale framework that integrates transfer learning and deep learning for urban land cover mapping. The framework utilizes pre-trained deep learning models, trained on diverse urban datasets, as a foundation for classification. These models are then finetuned using transfer learning techniques on smaller urban datasets, tailoring them to the specific characteristics of the target urban context. To evaluate the effectiveness and feasibility of the proposed framework, extensive evaluations are conducted across different cities and years. Performance metrics such as accuracy and dice score are employed to assess the framework's classification capabilities. The results of this study contribute to advancing the field of urban classification by demonstrating the effectiveness and feasibility of the cross-scale framework. By combining transfer learning and deep learning, the framework improves classification accuracy, efficiency, and scalability in urban land cover mapping tasks. Leveraging the strengths of transfer learning and deep learning holds great promise for accurate and efficient urban land cover mapping, providing valuable insights for urban planning and resource management decision-making.