Bridging the Gap: Evaluating Transfer Learning Strategies for Block-Level Urban Land Use Classification on Regional Empirical High-Resolution Remote Sensing Imagery

Transfer learning, particularly through pre-training and fine-tuning strategies, is widely used to address the substantial training sample requirement of deep learning methods for urban land use classification using high spatial resolution remote sensing imagery. However, most studies utilize natural images as the source domain, which still presents significant differences compared to remote sensing images. Furthermore, most existing studies focus primarily on the accuracy of different strategies, often overlooking that they are tested on experimental non-geographical grid patch datasets instead of regional empirical datasets. This oversight raises concerns about the replicability of their high accuracy in practical urban land use settings. In this study, we employed a general object-oriented deep learning framework for remote sensing classification of urban land use, using Beijing as the empirical region, to assess the effectiveness of different pre-training and fine-tuning methods on practical geospatial data. Our findings reveal that while accuracy on non-geographical grid scale datasets reached 88%, it significantly dropped to 67% when applied to the empirical data in Beijing region. By incorporating a two-stage pre-training approach using natural image dataset as the source domain and integrating a remote sensing image dataset as an intermediate domain, this technique improves regional classification accuracy by 3.78%. Our results indicate four factors contributing to the accuracy discrepancy: sample selection bias, mixed-use within blocks, imbalanced land use class proportions, and spatial heterogeneity.The study underscores the importance of regional empirical experiments and offers insights into mitigation methods for accuracy discrepancies in the transfer learning application to remote sensing.