Abstract: In recent years, convolutional neural networks (CNNs) have emerged as a core technique for semantic segmentation of remote sensing images owing to their powerful capabilities in automatic feature extraction and deep modeling. This study systematically reviews the research progress in this field and determines the technical bottlenecks and development directions. First, based on differences in the core technical architecture, the CNN-based methods for semantic segmentation of remote sensing images are classified into five categories: sliding window-or candidate region-based methods, encoder-decoder-based methods, dilated convolution-based methods, methods integrating attention mechanisms, and lightweight methods. These methods are detailed from the perspective of basic philosophy, technical features, and advantages and disadvantages. Then, through comparative experiments on multiple datasets, the segmentation accuracy and efficiency of various CNN models are quantitatively analyzed. Accordingly, the mechanisms underlying the coupled influence of model structure design, key parameter configuration, and dataset characteristics on segmentation results are revealed. Meanwhile, it is confirmed that the same model faces challenges in maintaining consistent performance and advantages across different datasets and various target segmentation tasks. Finally, this review presents core challenges confronting current research, including the inherent contradiction between accuracy and efficiency, high dependency on large-scale high-quality labeled samples, insufficient integration of geoscientific prior knowledge, and the limited spatiotemporal generalization ability of models. To address these challenges, this study puts forward future development paths. Specifically, it is recommended to explore semi-supervised/unsupervised learning, integrate the technical advantages of multiple models and complementary multi-modal data, and establish a network architecture characterized by the collaborative optimization of lightweight design and high accuracy. Furthermore, it is necessary to enhance the organic integration of geoscientific prior knowledge and deep models. This effort will help enhance the generalization ability and interpretability of models.