The goal of historic image classification using signature extension techniques is of widespread interest for remote sensing applications. Historic ground reference data for classifier training is frequently unavailable, necessitating unsupervised transfer learning techniques. Environmental and data variability over time create variable spectral target classes, creating a challenge for knowledge transfer. There is a need for robust classifiers that account for data with high variability and non-normal distributions, and can be successfully applied to unlabeled data; ensemble classification trees are one solution to this problem. With the goal of detecting submerged aquatic vegetation, we trained an ensemble classifier with one collection of 48 imaging spectroscopy flightlines from a single year (2008). We then tested its performance when applied to four years (2004-2007) of historic imaging spectroscopy of the same size over the same area. We validated the resulting classifications with corresponding historic ground reference data. Knowledge transfer success was varied, depending on which image year dataset was used to train the classifier. The 2008-trained classifier had the most stable performance when applied to historic image datasets: overall accuracies ranged from 78.8% to 85.9%. Detection of submerged aquatic vegetation was limited most by the percent cover of the canopy. Water column depth above the plant canopy did not have an effect on detection. The classifier was successful because the training dataset encompassed the range of variability of the study area and the historic datasets.