Lower bounds in classification for feature and algorithm selection

The objective of this paper is to study recent advancements in estimation lower bound classification results. These lower bounds are estimated for a given set of features, targets, Signal to Noise Ratios (SNRs), and representative clutter environments. The motivation of this work comes from the desire to know the best achievable classification results for a given set of features at a range of SNRs and sensor data. This will assist the end user and classifier designer to select features that maximize the theoretical classification performance (i.e. minimize the classification errors in the confusion matrix tables). It will also assist in selecting the suitable classification algorithms approaching the lower theoretical classification bounds. The theoretical bounds used in this paper in our experimental examples are based on the Bayesian approach. However, other bounds are also reviewed. These results can be applied for selecting features and classifiers for earth and deep space observations and surveillance.