The state of charge (SOC) estimation for lithium-ion batteries is highly sensitive to temperature. Traditional methods based on equivalent circuit models require offline open circuit voltage (OCV) testing at different temperatures to obtain OCV-SOC curves, and the process is extremely time-consuming. Meanwhile, large data-based methods necessitate collecting large amounts of feature data at various temperatures for model training. Both approaches are difficult to apply for online estimation. A new online SOC estimation method across different temperatures was propesed. An incremental support vector machine (ISVM) models the terminal voltage, while online collection of mini-batch data enables dynamic model parameter updates. And an adaptive unscented Kalman filter (AUKF) ensures closed-loop estimation. Tests under federal urban driving schedule (FUDS) at 0 ℃, 25 ℃, and 45 ℃ show that the method achieves SOC estimation errors within 0.02%.Compared to conventional methods, the mean absolute error and root mean square error reduce by 1.09 percentage points and 1.10 percentage points, respectively. The method also exhibits strong robustness to initial SOC values.