The state of health (SOH) of lithium-ion batteries (LIBs) is critical for power storage systems. The present data-driven approaches for SOH estimation are restricted by the randomness of the discharging process, which makes it challenging to extract health features (HFs). Multiple HFs such as the peak value and peak position of incremental capacity (IC) curves, the area of voltage curve in a specific voltage interval, and the current change rate are selected by using incremental capacity analysis (ICA) and voltage-current analysis (VCA) during the charging process, and then the correlation between the selected HFs and SOH is calculated by Pearson coefficient. Four HFs with the highest correlation are selected and finally validated using Gaussian process regression (GPR) under two training modes for SOH estimation. This paper uses NASA dataset to verify the accuracy of the estimation method and uses an experimental dataset to verify the validity of the method under a low temperature for aging. The results demonstrate that the HFs based on the charging process with GPR model are highly applicable. It can accurately estimate SOH of LIBs operating at different temperatures using only charging data and exhibits excellent generalization capability.