MARTA GANs: Unsupervised Representation Learning for Remote Sensing Image Classification

Abstract: With the development of deep learning, supervised learning has frequently been adopted to classify remotely sensed images using convolutional networks (CNNs). However, due to the limited amount of labeled data available, supervised learning is often difficult to carry out. Therefore, we proposed an unsupervised model called multiple-layer feature-matching generative adversarial networks (MARTA GANs) to learn a representation using only unlabeled data. MARTA GANs consists of both a generative model $G$ and a discriminative model $D$. We treat $D$ as a feature extractor. To fit the complex properties of remote sensing data, we use a fusion layer to merge the mid-level and global features. $G$ can produce numerous images that are similar to the training data; therefore, $D$ can learn better representations of remotely sensed images using the training data provided by $G$. The classification results on two widely used remote sensing image databases show that the proposed method significantly improves the classification performance compared with other state-of-the-art methods.

An Expert Overview of "MARTA GANs: Unsupervised Representation Learning for Remote Sensing Image Classification"

The paper, "MARTA GANs: Unsupervised Representation Learning for Remote Sensing Image Classification," addresses the challenges inherent in classifying high-resolution satellite images by leveraging an unsupervised learning approach via Generative Adversarial Networks (GANs). Specifically, the proposed multiple-layer feature-matching GANs (MARTA GANs) methodology seeks to overcome the limitations imposed by the scarcity of labeled data in remote sensing image classification.

Background and Motivation

Remote sensing image classification is instrumental for applications in land-resource management and urban planning. Traditional methods, such as the bag of visual words (BoVW) and spatial pyramid matching (SPM), rely on hand-crafted features which are typically labor-intensive to produce and require domain expertise. Although deep learning techniques using Convolutional Neural Networks (CNNs) have shown promising results, they necessitate extensive labeled datasets which are often unavailable in the remote sensing domain.

To address these issues, the authors introduce MARTA GANs, a novel unsupervised learning model, which consists of a generative model $G$ and a discriminative model $D$. MARTA GANs aim to learn effective representations from unlabeled data by generating images similar to the training dataset, thus enabling $D$ to extract more discriminative features. The paper reports enhanced classification results on UC Merced Land Use and Brazilian Coffee Scenes datasets compared to existing state-of-the-art methods.

Methodology

MARTA GANs expand on the DCGAN architecture by introducing several refinements:

• Image Resolution: The approach generates images up to a resolution of $256 \times 256$, higher than the $64 \times 64$ produced by standard DCGANs, achieved by incorporating additional deconvolutional layers.
• Feature Extraction: The discriminator acts as a feature extractor through a multi-feature layer, aggregating mid- and high-level information.
• Loss Functions: The model employs perceptual loss and feature matching loss to improve the accuracy and realism of synthesized images.

The methodology involves training the discriminator to differentiate real images from synthesized ones while simultaneously training the generator to produce images that deceive the discriminator, employing a minimax game strategy.

Results

The experimental results on both datasets demonstrate MARTA GAN's superior performance in unsupervised scene classification compared to DCGANs and other unsupervised methods. Specifically, MARTA GANs achieve an overall classification accuracy of $94.86\%$ on the UC Merced dataset and $89.86\%$ on the Brazilian Coffee Scenes dataset, showcasing its ability to extract robust features in complex scenarios.

Implications and Future Work

This research underscores the potential of unsupervised learning models like MARTA GANs in domains constrained by limited labeled data availability. By harnessing GANs' capability to synthesize plausible training data, MARTA GANs lessen the dependency on extensive labeled datasets, thereby opening avenues for more cost-effective and scalable remote sensing applications.

Looking ahead, further developments may entail enhancing the quality of generated samples from remote sensing images or extending the classification framework via semi-supervised learning models to boost classification accuracy. Additionally, exploring MARTA GANs in varied remote sensing contexts could delineate its versatility and adaptability to different satellite image complexities.

In summary, MARTA GANs offer a compelling unsupervised alternative that paves the way for more advanced remote sensing classification methodologies, facilitating automated and precise image interpretation in vast and varied datasets.