This paper develops a novel just-in-time learning (JITL) based soft sensor for multimode processes. The involved multimode data sets are assumed to be non-Gaussian distributed and time varying. A supervised non-Gaussian latent structure (SNGLS) is introduced to model the relationship between predictor variables and quality variables. In order to handle the multimode process, a moving window approach is adopted, based on which a new similarity measure is proposed by integrating window confidence and between sample local similarity. The similarity between the current query sample and the data set in a specific window is quantified by the window confidence using the support vector data description (SVDD). On the basis of the data in the moving window, the SNGLS model is constructed and used to obtain the estimation of between-sample local similarity. The two similarities are integrated and used as sample weights, and a locally weighted structure is designed for key quality variable estimation. The performance of the developed method is demonstrated by application studies to the Tennessee Eastman (TE) process and a predecarburization absorption unit. It is shown that the proposed method outperformed competitive methods on the prediction accuracy of key quality variables.