Stacked Random Forests: More Accurate and Better Calibrated

Abstract

Stacked Random Forests (SRFs) sequentially apply multiple Random Forests (RFs) where each instance uses the estimate of the predecessor as additional input to further refine the prediction. They have been shown to improve the performance for semantic segmentation of Polarimetric Synthetic Aperture Radar (PolSAR) images. Both, RFs and SRFs, not only provide an estimate of the class label of a query sample, but instead make a probabilistic prediction, i.e. provide the full class posterior. The probabilistic predictions of RFs are known to be usually well calibrated (i.e. the predictions match the expected probability distributions of each class). This paper answers the question whether stacking leads to overfitting on the training data or decreases the calibration quality of RFs. Results indicate that neither is the case. Instead, classification accuracy steadily increases and then saturates quickly after only a few stacking levels. The predicted probabilities are generally well calibrated where calibration quality also increases slightly for higher stacking levels.