Improving Deep Forest via Patch-Based Pooling, Morphological Profiling, and Pseudo Labeling for Remote Sensing Image Classification

Deep forest (DF), an alternative to neural networks (NNs)-based deep learning (DL), has gained increasing attentions in recent years. Despite its remarkable advantages, the original multigrained cascade forest (gcForest) is limited by the high time cost and memory requirement. To overcome this limitation, gcForest with confidence screening (gcForestCS) and feature screening (gcForestFS) were proposed with the proven improvements. But they were not comparatively studied for remote sensing (RS) image classification. Furthermore, gcForest, gcForestCS, and gcForestFS could be further improved by introducing patch-based pooling (PP), morphological profiling (MP), and pseudo labeling (PL) techniques. In this sense, DF algorithms are introduced and comparatively studied for hyperspectral and polarimetric synthetic aperture radar (PolSAR) image classification first. To further foster the classification performance from accurate, efficient, and effective feature abstraction viewpoints, improved versions of gcForest, gcForestCS, and gcForestFS, are proposed by adopting PP, MP, and PL techniques. To evaluate the performance of the introduced and proposed DF algorithms, six state-of-the-art spectral-spatial features aware NNs based DL algorithms are selected. Experimental results on three widely acknowledged hyperspectral and PolSAR benchmarks showed that: 1) gcForest, gcForestCS, and gcForestFS are also advanced algorithms for RS image classification; 2) mixed pooling with larger patch size set is always the best option in contrast with average, maximum, minimum, and median pooling strategies; and 3) positive improvements on gcForest, gcForestCS, and gcForestFS are clear using PP, MP, and PL techniques, and the best improvements can always be obtained by fused usage of PP and MP with PL features.