The performance of classification algorithms is heavily related to the quality of the training samples in supervised learning. Conventional statistical classifiers assume that data have a specific distribution. Such assumptions may not be valid for real world data. Additionally, enough training samples are required for every class to make a proper estimation of parameters to represent distribution functions. In general, there is a limited number of training samples in remote sensing. Therefore, classification algorithms should be robust with various types of training sample sets to achieve sufficient generalization performance. In this study, a new classification algorithm called border feature detection and adaptation (BFDA) is used to partition the feature space by taking into account some geometric considerations to support maximum margins between different class borders via some reference vectors called border features. The performance of the BFDA is related to the initialization of the border features during the border feature detection stage, and the input ordering of the training samples during the adaptation process. These dependencies cause relatively biased decisions. Therefore, consensual strategy with cross validation can be applied to improve the generalization performance. The resulting process is called consensual BFDA (C-BFDA).