Abstract
Domain adaptation aims to enhance the feature transferability of a model across different domains with feature distribution differences, which has been widely explored in many computer vision tasks such as semantic segmentation and object detection, but has not been fully studied in hyperspectral image (HSI) classification task. Compared with the natural image-based domain adaptation, HSI-based domain adaptation still faces two main challenges: 1) Due to the strong spectral variability of HSIs, it is difficult to extract discriminative and domain-invariant features from different domains, resulting in the misalignment of cross-domain features; 2) Class-wise (or fine-grained) spectral feature inconsistency between domains also inevitably degrades the classification accuracy. To address these issues, in this paper, we propose a novel coarse-to-fine joint distribution alignment framework for cross-domain classification of HSIs. Specifically, the training samples from source and target domains are first fed into a coupled variational auto-encoders (VAE) module, which is composed of two well-designed VAEs equipped with mutual information metric to learn high-level domain-invariant representations in a shared latent space, so that the network can learn a coarse-grained source-target feature consistency. Furthermore, to alleviate the class-wise inter-domain feature inconsistency, a joint distribution alignment (JDA) module is constructed to perform a fine-grained cross-domain alignment by matching the joint probability distributions between the source and target domains through adversarial learning. Extensive experiments on both simulated and real hyperspectral datasets demonstrate the superiority of the proposed method in comparison with several conventional and state-of-the-art methods.
Highlights
HYPERSPECTRAL images (HSIs), which are captured by the hyperspectral remote sensors, have become one of the most important data types in the field of remote sensing
We focus on the study of the Unsupervised domain adaptation (UDA) for the HSI classification task
Based on the well-aligned marginal distribution of features that can achieve a coarse-grained cross-domain feature consistency, we propose a joint distribution alignment (JDA) module that can explicitly align the joint distribution of the latent codes and labels from the source or target domain, ensuring class-wise fine-grained feature consistency
Summary
HYPERSPECTRAL images (HSIs), which are captured by the hyperspectral remote sensors, have become one of the most important data types in the field of remote sensing. Unlike the traditional optical imaging sensors, HSIs are usually composed of several hundreds of spectral data channels collected in the same scene, and present rich spectral information. HSI classification aims to assign a land-cover class label to each pixel within a hyperspectral imagery, and it. This manuscript was first submitted on Aug. 20, 2021 for review. As a result, adapting an existing model to new images without relying on manually labeling information is necessary and significant for practical applications. When the labeled samples in the HSIs that need to be classified are not existent (i.e. the target dataset), we aim to design a network architecture for cross-scene HSI classification where the model trained on a source HSI with sufficient labels can classify the unlabeled target dataset well
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
More From: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.