In the context of rapid urbanization, climate change, and increasing demand for higher living standards, urban development has gradually shifted from outward expansion to internal renewal. Planning and optimizing urban spaces are now central to enhancing residents' quality of life and promoting sustainable growth. However, traditional urban evaluation methods, often based on expert assessments and quantitative indicators, tend to overlook the diverse and subjective experiences of citizens. As social media platforms increasingly serve as spaces for public expression, geotagged user-generated content offers a valuable yet underutilized resource for capturing dynamic urban perceptions. Meanwhile, the rise of large language models (LLMs) provides powerful tools to process and extract insights from these massive, unstructured datasets. This study focuses on Harbin, a cold-climate tourism city in China, to develop a multi-dimensional urban perception evaluation framework based on geotagged social media comments. First, we collected a dataset of geotagged posts from the Weibo platform between 2020 and 2025, followed by rigorous data cleaning to ensure relevance and eliminate noise. Next, leveraging LLMs, we categorized the comments into four key perception dimensions: natural environment, public space, sensory experience, and cultural identity. Representative keywords were extracted to enrich semantic understanding within each dimension. Then, sentiment analysis was conducted to compute satisfaction scores for each dimension, revealing the emotional valence of urban experiences. Finally, spatial visualization techniques, including heatmaps and clustering analysis via ArcGIS, were employed to map the distribution of sentiments across Harbin, thereby identifying areas with notable positive and negative perceptions. To validate the robustness of the LLM-based analysis, we conducted a comparative study involving expert-labeled data. The automated categorization and sentiment scoring achieved over 85% consistency with human annotations, significantly outperforming traditional LDA-based models in handling nuanced and conflicting expressions. Our findings indicate that public dissatisfaction primarily centers around challenges of climate adaptability, such as extreme winter temperatures and slippery, icy roads, whereas cultural landmarks—particularly Russian-style historic architecture—garner strong positive evaluations. Temporal analysis further reveals significant seasonal variation: perceptions of ice and snow attractions are notably more positive during winter months compared to non-winter periods, suggesting the importance of seasonal adaptation in urban design. Based on these insights, we propose adaptive strategies such as establishing windbreak corridors at key winter attractions like the Ice and Snow World, and implementing time-based traffic flow regulations on Central Street to enhance pedestrian safety and comfort. This research demonstrates the potential of a "low-code, high-semantic" analytical paradigm to integrate citizen feedback into urban renewal processes. By generating spatially associated perception maps, the study offers actionable guidance for sustainable urban planning, especially in cold-climate cities seeking to balance tourism development with climate resilience and quality of life improvements. Keywords: Social media; Urban perception evaluation; Large language models (LLMs); Spatial planning; Sustainable development; Cold-region cities; Harbin