Ensemble learning soft sensors have been widely used to estimate key quality parameters in the process industry. However, the conventional ensemble modeling methods are often limited to mining the temporal relationships between samples for building the base models while ignoring the spatial relationships between samples. This may lead to problems such as insufficient local state mining of the process and insufficient diversity among base models. In addition, traditional soft sensor methods cannot effectively deal with the timevarying characteristics of the process due to the lack of adaptive mechanisms, which leads to the degradation of the model performance. Therefore, an ensemble adaptive soft sensor method based on the spatio-temporal local learning (STLL) is proposed. Firstly, the method mines the temporal and spatial relationships of process data, and the redundant states are removed by using statistical hypothesis testing. Then, a set of diverse spatial-temporal local Gaussian mixture regression models ( GMR) is formulated. Then, the local prediction results are combined adaptively based on an online selective ensemble strategy. Besides, a dual-updating strategy is proposed for alleviating the model performance degradation. Compared to the non-adaptive global GMR, temporal local learning based ensemble GMR, spatial local learning based ensemble GMR, experimental results show that the prediction accuracy of the proposed STLL approach is improved by 70. 3% , 14. 9% , and 27. 8% in an industrial chlortetracycline fermentation process, while it is improved by 31. 9% , 21. 2% , and 19. 3% in an industrial debutanizer process.