Fig. 4.

Flow chart of training CNN_5layers and further improving predictive performance under 5-fold cross-validation. The datasets were randomly and equally separated into five subgroups, and alternately four subgroups were used as the training set to train the models. The remaining subgroup was used to test the generalization capacity of the trained models. We combined trained CNN_5layers output with extra features to continue training simple linear regression models. Finally, the remaining subgroup was again used to test improved performances