Thermal sensation vote (TSV) is a key indicator in thermal comfort research due to its integration of subjective perception with environmental conditions. Existing methods of estimating TSV rely on linear estimation, overlooking the intrinsically nonlinear influence of human thermal perception. To address such limitations, this study aims to systematically investigate how indoor meteorological variables shape students’ thermal sensation by coupling measurements, questionnaires, and explainable machine learning (ML). Environmental parameters, including air temperature (T_a), globe temperature (T_g), relative humidity (RH), and air velocity (V_a), together with students’ thermal sensation votes (TSV), were collected across transitional, summer, and winter conditions. Based on the coupled subjective–objective dataset, a random forest model was developed to quantify TSV under realistic classroom conditions. The model demonstrates good performance across seasonal scenarios (R² = 0.992, RMSE = 0.094), indicating its effectiveness in capturing nonlinear relationships between subjective thermal sensation and environmental variables. Model interpretation was supported through SHapley Additive exPlanations analysis, enabling identification of the most dominant driver as T_g, followed by T_a, RH and V_a. Scenario-based analysis further shows that a 1 ℃ increase in T_g is associated with an increase in TSV of approximately 0.35 to 0.50, compared with only 0.10 to 0.18 for an equivalent increase in T_a, highlighting the substantially stronger influence of radiant environment on perceived comfort. Overall, this study provides a robust, data-driven framework for analyzing thermal sensation in secondary school classrooms and offers practical support for performance-oriented indoor environmental control and design in hot–humid climates.