Spectral-Spatial Neural Networks and Probabilistic Graph Models for Hyperspectral Image Classification

Abstract

Pixel-wise hyperspectral image (HSI) classification has been actively studied since it shares similar characteristics with related computer vision tasks, including image classification, object detection, and semantic segmentation, but also possesses inherent differences. The research surrounding HSI classification sheds light on an approach to bridge computer vision and remote sensing. Modern deep neural networks dominate and repeatedly set new records in all image recognition challenges, largely due to their excellence in extracting discriminative features through multi-layer nonlinear transformation. However, three challenges hinder the direct adoption of convolutional neural networks (CNNs) for HSI classification. First, typical HSIs contain hundreds of spectral channels that encode abundant pixel-wise spectral information, leading to the curse of dimensionality. Second, HSIs usually have relatively small numbers of annotated pixels for training along with large numbers of unlabeled pixels, resulting in the problem of generalization. Third, the scarcity of annotations and the complexity of HSI data induce noisy classification maps, which are a common issue in various types of remotely sensed data interpretation.

Recent studies show that taking the data attributes into the designing of fundamental components of deep neural networks can improve their representational capacity and then facilitates these models to achieve better recognition performance. To the best of our knowledge, no research has exploited this finding or proposed corresponding models for supervised HSI classification given enough labeled HSI data. In cases of limited labeled HSI samples for training, conditional random fields (CRFs) are an effective graph model to impose data-agnostic constraints upon the intermediate outputs of trained discriminators. Although CRFs have been widely used to enhance HSI classification performance, the integration of deep learning and probabilistic graph models in the framework of semi-supervised learning remains an open question.

To this end, this thesis presents supervised spectral-spatial residual networks (SSRNs) and semi-supervised generative adversarial network (GAN) -based models that account for the characteristics of HSIs and make three main contributions. First, spectral and spatial convolution layers are introduced to learn representative HSI features for supervised learning models. Second, generative adversarial networks (GANs) composed of spectral/spatial convolution and transposed-convolution layers are proposed to take advantage of adversarial training using limited amounts of labeled data for semi-supervised learning. Third, fully-connected CRFs are adopted to impose smoothness constraints on the predictions of the trained discriminators of GANs to enhance HSI classification performance. Empirical evidence acquired by experimental comparison to state-of-the-art models validates the effectiveness and generalizability of SSRN, SS-GAN, and GAN-CRF models.