Domain adaptation framework with ensemble of fuzzy rules-based ELMs for remote-sensing image classification.

The domain adaptation (DA) transfer learning technique can accurately classify land cover in remote-sensing (RS) images, even with a small number of labeled samples. However, traditional architectures and various neural network and deep learning modifications make DA models difficult to learn quickly and computationally demanding. Extreme learning machine (ELM)-based DA models have shown potential for quick learning and improved generalization capabilities. Still, they are also known for their unstable and over-fitting characteristics and the ambiguity of the data sets. To address these difficulties, this article proposes two contributions. First, a modified fuzzy-rule-based ELM model (MFR-ELM) is created within a data-dependent platform to address the imprecision of the RS image land-cover classes and small sample size. Second, an ensemble of fuzzy rule-based ELM networks is proposed in the DA framework to address individual networks' instability and weak learning properties. The MFR-ELM and other fuzzy rule-based ELM designs are combined in the ensemble architecture for mutual benefit. The suggested model is proven to be superior to competing approaches for categorizing multispectral RS images, as supported by various performance assessment indices. [ABSTRACT FROM AUTHOR]