Ensemble learning has become a widely used soft sensor modeling framework. However, the establishment of high-performance ensemble learning soft sensor models still encounters many challenges such as improper feature selection, insufficient diversity of base models, and poor base model estimation performance. Therefore, a selective ensemble of stacked autoencoder based Gaussian process regression (SESAEGPR) is proposed for soft sensor modeling. By fully utilizing the advantages of deep learning in feature extraction, the SESAEGPR method first builds a set of diverse stacked autoencoder(SAE) networks and then establishes a set of Gaussian process regression (GPR) models based on the already extracted latent features. Then, a two-stage ensemble pruning is performed. The first is achieved based on the model performance improvement, and the evolutionary multi-objective optimization approach is used for the second. Ensemble pruning enables the reduction of ensemble model complexity while maintaining or even further improving the ensemble estimation performance. Finally, a PLS Stacking ensemble mechanism is employed to achieve the combination of the selected base models. The proposed method performs significantly better than the traditional global and fully integrated soft sensing methods, and its effectiveness and superiority have been verified through the penicillin fermentation process and the Tennessee Eastman chemical process.