Automated colon cancer detection helps get rid of the slow and laborious process of manual examination of histopathological tissue specimens using microscope, and provides a reliable second opinion to the histopathologists. Therefore, automated colon cancer detection has been the focus of research community in the past two decades, and researchers have proposed various automatic colon cancer detection systems. Most of the existing colon cancer detection systems extract features and then construct single classifier to perform classification. However, the small sample size problem and especially the noise in neuroimaging data makes it challenging to achieve good classification results by training only one global classifier. In this work, we propose a local patch based ensemble method instead of building a single global classifier. The proposed method builds multiple individual weak classifiers based on the different subsets of local patches, and then combines the output of weak classifiers for more accurate and robust classification. In particular, Haralick and Local Binary Patterns (LBP) features are extracted from pre-processed colon biopsy images, and images are partitioned into smaller fixed size patches. Features of the random subsets of patches are combined to train weak classifier. Several kernels of SVM such as linear, RBF and sigmoid are used as weak classifiers. Later, the output of various weak classifiers is combined to get the final classification results. The proposed subspace ensemble classification method yields better results compared to one global classifier in all the three cases (linear, RBF, sigmoid) in terms of various performance measures such as accuracy, sensitivity, specificity, receiver operating characteristics (ROC) curves, Matthew's correlation coefficient, and Area Under the Curve (AUC), however, classification performance is slightly better for ensemble of RBF kernel. The proposed method has also yielded better performance compared to existing techniques on a colon cancer dataset of 174 subjects.