Empirical Orthogonal Function (EOF) analysis is used in this study to generate main eigenvector fields of historical temperature for the China Seas (here referring to Chinese marine territories) and adjacent waters from 1930 to 2002 (510 143 profiles). A good temperature profile is reconstructed based on several subsurface in situ temperature observations and the thermocline was estimated using the model. The results show that: 1) For the study area, the former four principal components can explain 95% of the overall variance, and the vertical distribution of temperature is most stable using the in situ temperature observations near the surface. 2) The model verifications based on the observed CTD data from the East China Sea (ECS), South China Sea (SCS) and the areas around Taiwan Island show that the reconstructed profiles have high correlation with the observed ones with the confidence level >95%, especially to describe the characteristics of the thermocline well. The average errors between the reconstructed and observed profiles in these three areas are 0.69°C, 0.52°C and 1.18°C respectively. It also shows the model RMS error is less than or close to the climatological error. The statistical model can be used to well estimate the temperature profile vertical structure. 3) Comparing the thermocline characteristics between the reconstructed and observed profiles, the results in the ECS show that the average absolute errors are 1.5m, 1.4 m and 0.17°C/m, and the average relative errors are 24.7%, 8.9% and 22.6% for the upper, lower thermocline boundaries and the gradient, respectively. Although the relative errors are obvious, the absolute error is small. In the SCS, the average absolute errors are 4.1 m, 27.7 m and 0.007°C/m, and the average relative errors are 16.1%, 16.8% and 9.5% for the upper, lower thermocline boundaries and the gradient, respectively. The average relative errors are all <20%. Although the average absolute error of the lower thermocline boundary is considerable, but contrast to the spatial scale of average depth of the lower thermocline boundary (165 m), the average relative error is small (16.8%). Therefore the model can be used to well estimate the thermocline. Empirical Orthogonal Function (EOF) analysis is used in this study to generate main eigenvector fields of historical temperature for the China Seas (here referring to Chinese marine territories) and adjacent waters from 1930 to 2002 (510 143 profiles). A good temperature profile is reconstructed based on several subsurface in situ temperature observations and the thermocline was estimated using the model. The results show that: 1) For the study area, the former four principal components can explain 95% of the overall variance, and the vertical distribution of temperature is most stable using the in situ temperature observations near the surface. 2) The model verifications based on the observed CTD data from the East China Sea (ECS), South China Sea (SCS) and the areas around Taiwan Island show that the reconstructed profiles have high correlation with the observed ones with the confidence level> 95%, especially to describe the characteristics of the thermocline well. The average errors between the reconstructed and observed profiles in these three areas are 0.69 ° C, 0.52 ° C and 1.18 ° C respectively. It also shows the model RMS error is less than or close to the climatological error. The statistical model can be used to well estimate the temperature profile vertical structure. 3) Comparing the thermocline characteristics between the reconstructed and observed profiles, the results in the ECS show that the average absolute errors are 1.5m, 1.4 m and 0.17 ° C / m, and the average relative errors are 24.7 In the SCS, the average absolute errors are 4.1 m, 27.7 m and 0.007 ° C / m, and the average relative errors are 16.1%, 16.8% and 9.5% for the upper, lower thermocline boundaries and the gradient, respectively. The average relative errors are all <20%. Although the average absolute error of the lower thermocline b oundary is considerable, but contrast to the spatial scale of average depth of the lower thermocline boundary (165 m), the average relative error is small (16.8%). Therefore the model can be used to well estimate the thermocline.