Rapid inter-day temperature fluctuations pose increasing risks to coupled human–environment systems, yet the spatiotemporal dynamics of public sensitivity to such variability remain poorly quantified. This study develops a geospatial–computational framework integrating deep learning-based social sensing with explainable machine learning. Analyzing approximately 1.9 million geotagged temperature-related posts from 367 Chinese cities across three representative years (2017, 2020, and 2022), we constructed a city-scale sensitivity index to measure residents' responsiveness to thermal shocks. Results reveal a fundamental asymmetry, with perception sensitivity to cooling (0.114%/°C) being 2.3 times higher than to warming (0.049%/°C) on average, though this gap narrows under extreme temperatures (peaking at 0.286%/°C and 0.296%/°C, respectively). Sensitivity shows marked spatial heterogeneity and regional clustering, with high-sensitivity hotspots concentrated in southwestern China and low-sensitivity coldspots in northwestern and northeastern regions. Furthermore, factor analysis reveals that climatic baselines (e.g., inter-day temperature variability, precipitation), environmental characteristics (e.g., elevation), and socioeconomic factors (e.g., population density) significantly modulate perception sensitivity. These findings highlight the urban amplification effect and validate social media as a real-time sensor for human thermal stress. Our approach supports spatially targeted adaptation planning and the integration of subjective perception into urban early warning systems.