Enriching Earth Observation labeled data with Quantum Conditioned Diffusion Models

TL;DR

The paper tackles the scarcity of labeled Earth Observation data and the computational demands of diffusion models by introducing QCU-Net, a class-conditioned quantum diffusion framework that integrates quanvolutional feature extraction and a variational quantum circuit at the bottleneck. Through extensive experiments on EuroSAT RGB, it shows substantial improvements in image realism (FID/KID) and semantic fidelity over classical baselines, with ablations highlighting the importance of early-encoder quantum placement and HQConv design. The approach enables high-quality, labeled EO data generation for data augmentation and downstream tasks, marking a significant step toward quantum-enhanced remote sensing. The work also discusses practical considerations, including training stability, parameter efficiency, and avenues for future multi-sensor EO applications and hardware-aware scaling.

Abstract

The rapid adoption of diffusion models (DMs) in the Earth Observation (EO) domain has unlocked new generative capabilities aimed at producing new samples, whose statistical properties closely match real imagery, for tasks such as synthesizing missing data, augmenting scarce labeled datasets, and improving image reconstruction. This is particularly relevant in EO, where labeled data are often costly to obtain and limited in availability. However, classical DMs still face significant computational limitations, requiring hundreds to thousands of inference steps, as well as difficulties in capturing the intricate spatial and spectral correlations characteristic of EO data. Recent research in Quantum Machine Learning (QML), including initial attempts of Quantum Generative Models, offers a fundamentally different approach to overcome these challenges. Motivated by these considerations, we introduce the Quanvolutional Conditioned U-Net (QCU-Net), a hybrid quantum--classical architecture that applies quantum operations within a conditioned diffusion framework using a novel quanvolutional feature-extraction approach, for generating synthetic labeled EO imagery. Extensive experiments on the EuroSAT RGB dataset demonstrate that our QCU-Net achieves superior results. Notably, it reduces the Fréchet Inception Distance by 64%, lowers the Kernel Inception Distance by 76%, and yields higher semantic accuracy. Ablation studies further reveal that strategically positioning quantum layers and employing entangling variational circuits enhance model performance and convergence. This work represents the first successful adaptation of class-conditioned quantum diffusion modeling in the EO domain, paving the way for quantum-enhanced remote sensing imagery synthesis.

Figures (5)

• Figure 1: Class-conditioned Quantum Diffusion Model for EO image generation.Top: the forward diffusion process $q(x_t \mid x_{t-1})$ (in blue) progressively corrupts a clean RGB satellite patch $x_0$ into a Gaussian noise $x_T$. The learned reverse process $p_{\theta}(x_{t-1} \mid x_t, y)$ (in purple) then reconstructs the image step-by-step while being guided by the land-cover label $y$, enabling class-specific generation (e.g., residential, forest, river, crop). Bottom: the UNet denoiser integrates both classical components (Convolutional / ResNet / Attention blocks) and two quantum-enhanced residual variants. The QuanResNet Block (in red) replaces the first convolution with a quanvolutional quantum filter applied to local spatial patches. While the QResNet Block (in purple), placed at the bottleneck of the process, replaces both convolutions with a Variational Quantum Circuit (VQC) acting on $2\times2\times3$ feature blocks and applied to a fraction $\rho$ of channels.
• Figure 2: Hybrid quantum--classical ResNet. Top: comparison between the classical ResNet block and the QuanResNet block, in which one convolutional layer is replaced by a Quanvolutional (quantum) layer. Bottom: schematic of the HQConvAnsatz variational quantum circuit: angle encoding of feature maps, parameterized rotations, entangling gates, and measurements to generate quantum-enhanced features.
• Figure 3: Qualitative comparison between ground truth and generated images on the EuroSAT dataset. Columns correspond (left to right) to the ten land-cover classes: Annual Crop, Forest, Herbaceous Vegetation, Highway, Industrial, Pasture, Permanent Crop, Residential, River, and Sea Lake. Rows show ground truth reference images (first row), outputs of the classical model (second row), and outputs of the proposed quantum model (QCU-Net) (third row).
• Figure 4: Convergence trends over training iterations for different models and evaluation metrics (FID, KID, IS, and Accuracy). Each plot shows the evolution of performance across iterations for the classical baseline, the QVCU-Net and the proposed QCU-Net. In addition to FID (shown in the main text), similar trends are evident for the other metrics. Notably, qCU-Net not only achieves superior scores but also demonstrates a significantly faster convergence compared to the other approaches.
• Figure 5: Per-channel intensity distributions (densities; common bins in $[0,1]$) for AnnualCrop (top) and River (bottom). Vertical dashed and dotted lines denote mean and median. We compare the distribution for the GT, Classic model and the QCU-Net model.