Abstract
Earth Observation (EO) satellites generate huge volumes of high-resolution multi-modal data, allowing key applications such as land-cover mapping, crop identification, disaster response and environment monitoring. However, transforming this data into actionable information remains computationally demanding and at the same time, high-quality pixel-level annotations required for supervised learning are expensive and often unavailable at scale. Addressing both challenges, computational efficiency and limited supervision requires fundamentally new approaches to model design. In this work, we present a parameter-efficient hybrid Quantum Machine Learning (QML) architecture designed for downstream tasks. We develop a framework, that integrates quantum convolutional layers within an encoder-decoder backbone to enhance spectral-spatial representation learning while maintaining scalability. The quantum convolution layer employs small patch-based processing, where local spectral-spatial features are encoded into a compact multi-qubit parameterized quantum circuit. Through entangled transformations and measurement, the circuit produces non-linear feature mappings that capture cross-channel correlations beyond conventional convolutional filters. In addition, these quantum layers introduce strong representational capacity while using significantly fewer trainable parameters than classical deep learning architectures.
To address the challenges of high-resolution annotations, we incorporate a pseudo-label guidance strategy. A self-supervised reconstruction module learns intrinsic spectral-spatial distributions directly from the input imagery, while an auxiliary branch refines low-resolution labels into high-resolution pseudo-labels that guide segmentation. This dual-path training enforces consistency between reconstruction and semantic prediction, improving robustness to inaccurate supervision. The resulting model achieves robust semantic segmentation performance with reduced parameter counts and improved generalization across heterogeneous landscapes. Its compact design makes it particularly relevant for emerging onboard AI models. Modern satellites increasingly require real-time or near-real-time inference to reduce downlink bandwidth, enable rapid response to natural hazards and support autonomous decision-making. Parameter-efficient hybrid quantum-classical models offer a pathway toward such onboard intelligence by reducing memory footprint and computational overhead while preserving expressive feature extraction. The proposed implementations rely on classical simulation of quantum circuits on NVIDIA A100D-4-40C GPU under CUDA Version: 12.4, Ubuntu 22.04.5 LTS (GNU/Linux 6.8.0-87-generic x86-64). We have used pennylane $0.38.0$, torch $2.6.0+cu124$, rasterio $1.4.3$, matplotlib $3.7.5$ and numpy $1.26.4$ to implement the proposed approach and reference methods. As quantum processors become available, hybrid QML models may further explored and enhance onboard processing efficiency. The ongoing work demonstrates that quantum-enabled convolutional architectures can serve as lightweight and significant feature extractors for Earth Observation data analysis. By including parameter efficiency, pseudo-label guidance and hybrid quantum-classical learning, we propose a framework that bridges advanced machine learning with next-generation autonomous Earth Observation systems.