Discrete Wavelet Transform as a Facilitator for Expressive Latent Space Representation in Variational Autoencoders in Satellite Imagery

2025-09-30

AI Summary

Weak latent-space representation capability of variational autoencoders (VAEs) in remote sensing imagery limits the performance of subsequent latent diffusion models (LDMs). To address this, we propose ExpDWT-VAE—a dual-branch VAE architecture integrating discrete wavelet transform (DWT). Specifically, its encoder jointly processes spatial-domain features via convolution and frequency-domain features via 2D Haar wavelet decomposition, enabling explicit multi-scale frequency modeling in the latent space. The decoder reconstructs images through inverse wavelet transform while mapping diagonal Gaussian latent variables to compact yet semantically rich representations. Experiments on the TerraFly satellite image dataset demonstrate that ExpDWT-VAE significantly improves latent representation quality: it achieves superior PSNR, SSIM, and FID scores over baseline VAEs, thereby enhancing LDM-generated outputs in terms of fine-grained detail fidelity and texture consistency.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Responsible Web: Data and user privacy-enhancing technologies for the Web

Abstract

Latent Diffusion Models (LDM), a subclass of diffusion models, mitigate the computational complexity of pixel-space diffusion by operating within a compressed latent space constructed by Variational Autoencoders (VAEs), demonstrating significant advantages in Remote Sensing (RS) applications. Though numerous studies enhancing LDMs have been conducted, investigations explicitly targeting improvements within the intrinsic latent space remain scarce. This paper proposes an innovative perspective, utilizing the Discrete Wavelet Transform (DWT) to enhance the VAE's latent space representation, designed for satellite imagery. The proposed method, ExpDWT-VAE, introduces dual branches: one processes spatial domain input through convolutional operations, while the other extracts and processes frequency-domain features via 2D Haar wavelet decomposition, convolutional operation, and inverse DWT reconstruction. These branches merge to create an integrated spatial-frequency representation, further refined through convolutional and diagonal Gaussian mapping into a robust latent representation. We utilize a new satellite imagery dataset housed by the TerraFly mapping system to validate our method. Experimental results across several performance metrics highlight the efficacy of the proposed method at enhancing latent space representation.

Problem

• Enhancing VAE latent space representation using wavelet transform
• Improving satellite imagery analysis through frequency-domain features
• Addressing limited research on intrinsic latent space enhancements

Innovation

• Uses Discrete Wavelet Transform for VAE enhancement
• Integrates spatial and frequency domains via dual branches
• Creates robust latent representation for satellite imagery