Domain Adaptation for Remote Sensing Using Deep Learning

Date

2020-01-16

Authors

Elshamli, Ahmed

Journal Title

Journal ISSN

Volume Title

Publisher

University of Guelph

Abstract

Traditional machine learning (ML) techniques are often employed to perform complex pattern recognition tasks for remote sensing images, such as land-use classification. In order to obtain acceptable classification results, these techniques require sufficient training data to be available for every particular image. Obtaining training samples is challenging, particularly for near real-time applications. Therefore, past knowledge must be utilized to overcome the lack of training data in the current regime. This challenge is known as domain adaptation (DA), in which the training data (source) and the test data (target) are sampled from different domains. DA is a challenging problem and the performance of the proposed techniques can be significantly affected by the type of problem, the nature of the data, and the type of data shift associated with the domains. Even though more data can be obtained from different sources, adapting these sources to obtain acceptable results is also a challenging task. This is especially true when the different domains contain a severely imbalanced class distribution. In this study, different techniques based on deep neural networks (DNNs) were developed and evaluated to solve the DA problem for remote sensing image classification in different settings. First, the single-source DA problem was addressed by finding invariant representations for both the source and the target. Denoising autoencoders (DAE) and domain-adversarial neural networks (DANN) were adopted to find these invariant representations. Results showed that both techniques were able to outperform traditional approaches, such as principal component analysis (PCA) and kernel PCA. Second, the multi-source domain adaptation for large-scale applications was addressed. A novel, efficient, scalable, yet simple, adaptive multi-source domain adaptation (AMDA) was developed to address this problem. AMDA was also capable of dealing more effectively with the imbalanced data distribution among the sources. Two techniques originally proposed for domain expansion were also extended to the task of multi-source domain adaptation. AMDA and the extended domain expansion techniques were implemented and evaluated on the LCZ classification problem. Despite its simplicity, AMDA was able to achieve more than a 12% improvement over the baseline.

Description

Keywords

Remote Sensing, Machine Learning, Land-use, Deep Learning, Image Classification, Domain Adaptation

Citation