A Massively Parallel Deep Rule-Based Ensemble Classifier for Remote Sensing Scenes

Abstract

In this Letter, we propose a new approach for remote sensing scene classification by creating an ensemble of the recently introduced massively parallel deep (fuzzy) rule-based (DRB) classifiers trained with different levels of spatial information separately. Each DRB classifier consists of a massively parallel set of human-interpretable, transparent 0-order fuzzy IF…THEN… rules with a prototype-based nature. The DRB classifier can self-organize "from scratch" and self-evolve its structure. By employing the pre-trained deep convolution neural network as the feature descriptor, the proposed DRB ensemble is able to exhibit human-level performance through a transparent and parallelizable training process. Numerical examples using benchmark dataset demonstrate the superior accuracy of the proposed approach together with human-interpretable fuzzy rules autonomously generated by the DRB classifier. Index Terms-deep learning, rule-based classifier, scene classification, fuzzy rules. I. INTRODUCTION EMOTE sensing scene classification aims to allocate the sub-regions of fine spatial resolution images to distinct land use categories, a goal which is of paramount importance for many applications, such as urban planning, land resource management, and environmental conservation [1]. At the same time, land use classification is recognized widely as a challenging task because the land use sub-regions are recognised implicitly through their high-level semantic function, where multiple low-level features or land cover classes can appear in one land use category, and identical land cover classes can be shared among different land use categories. These high-level semantics need to be exploited sufficiently using robust and accurate approaches for feature representation. Currently, deep learning (DL) neural networks (NN) have gained huge popularity amongst researchers as well as amongst the general public, quickly becoming the state-of-art approach in the remote sensing domain, in particular [2]. Several publications have reported very promising results using DL for

Toward Anthropomorphic Machine Learning

“…The DRB system is able to offer a self-organizing, self-adaptive, transparent, highly parallelizable rulebased architecture and learning algorithm with theoretically proven convergence. It has to be stressed that this new method is a general machine learning approach and is applicable to various classification and prediction problems with simple modifications, but in this paper, we will present the general concepts and principles focusing on image classification problems [11]- [13].…”

“…1. As one can see from the figure, the classifier is composed of the following components [11]- [13]. 1) Pre-processing block, which involves the widely used pre-processing techniques in the field of computer vision including: i) normalization, ii) scaling, iii) rotation, and iv) segmentation, etc.…”

“…Each massively parallel fuzzy rule is formulated around a (large) number of prototypes identified from the training data samples of a particular class connected by a local decisionmaker using "winner takes all" principle. An example of such fuzzy rule is given in expression (1) [11]- [13]:…”

“…2(c) as an illustration. Another anthropomorphic characteristic of the DRB system is its strong sense of situation awareness, which is closely linked with its prototype-based nature [11]- [13] as well. For the newly observed samples, the DRB system will make a comparison with the previously identified prototypes.…”

Deep Rule-Based Aerial Scene Classifier using High-Level Ensemble Feature Descriptor

“…As a recently introduced generic approach for image classification, deep rule-based (DRB) classifier [12], [13] is a powerful alterative to the DCNN models. The DRB approach expands the traditional fuzzy rule-based (FRB) systems with a massively parallel multi-layer structure that DCNNs benefit from [13].…”