The project was developed by the CIN Researcher Francesco Mauro

Project Summary

• A State-of-the-Art Review on Quantum Machine Learning for Earth Observation (QML4EO) and Quantum Graph Neural Networks (QGNNs).
• Investigated the quanvolutional operator to efficiently extract quantum features from Earth-observation imagery, benchmarking it on four case studies: land-use/land-cover, speckle filtering, coastal water quality, and building segmentation.
• Developed a quantum latent diffusion model for satellite image synthesis.
• The results of the collaboration have been published in IEEE peer-reviewed remote-sensing journals and conferences, underscoring its scientific relevance and impact.

Development tools

• Language & IDEs: Python and VS Code.
• Quantum-ML stack: PennyLane and Hybrid Quantum Model library.
• Classical DL stack: JAX, PyTorch and Tensorflow.
• Data sources: Copernicus Sentinel-2 L1C/L2A, Sentinel-1 GRD, Copernicus Marine Service, Φ-Sat-2 simulated data.
• Compute: Julich HPC cluster

Development Outputs

• The code related to the developed projects will be updated on my GitHub repository: https://github.com/francescomauro1998?tab=repositories.

Liist of publications developed during this collaboration:

• Mauro, F., Razzano, F., Di Stasio, P., Sebastianelli, A., Meoni, G., Schirinzi, G., ... & Ullo, S. L. (2025). Quantum-Enhanced Water Quality Monitoring: Exploiting ΦSat-2 Data with Quanvolution. IEEE Geoscience and Remote Sensing Letters.
• Sebastianelli, A., Mauro, F., Ciabatti, G., Spiller, D., Le Saux, B., Gamba, P., & Ullo, S. (2025). Quanv4eo: empowering earth observation by means of quanvolutional neural networks. IEEE Transactions on Geoscience and Remote Sensing.
• Mauro, F., De Falco, F., Ceschini, A., Meoni, G., Sebastianelli, A., Panella, M., Gamba, P., & Ullo, S. L. (2025). Advancing Earth Observation with Trainable Quanvolutional Neural Networks for classification tasks. Accepted for IGARSS 2025-2025 IEEE International Geoscience and Remote Sensing Symposium. IEEE.
• De Falco, F., Mauro, F., Ceschini, A., Sebastianelli, A., Gamba, P., Ullo, S. L., & Panella, M. (2025). Leveraging Quantum Latent Diffusion Models for Data Augmentation\\ on the EuroSAT Dataset. Accepted for IGARSS 2025-2025 IEEE International Geoscience and Remote Sensing Symposium. IEEE.
• Russo, L., Mauro, F., Memar, B., Sebastianelli, A., Ullo, S.L., & Gamba P. (2025). A Quantum-assisted Attention U-Net for Building Segmentation over Tunis using Sentinel-1 Data. Presented to IEEE Joint Urban Remote Sensing Event (JURSE) 2025. IEEE.
• Ceschini, A., Mauro, F., De Falco, F., Sebastianelli, A., Verdone, A., Rosato, A., ... & Ullo, S. L. (2024). From Graphs to Qubits: A Critical Review of Quantum Graph Neural Networks. arXiv preprint arXiv:2408.06524. (Submitted to IEEE Transaction on Pattaern Analysis and Machine Intelligence).
• Sebastianelli, A., Mauro, F., Delilbasic, A., Fan, F., Meoni, G., Cavallaro, G., Xiang, X., Gamba, P., & Ullo, S. L. (2925). Quantum Machine Learning for Earth Observation: A Review and Future Prospects. Submitted to IEEE Geoscience and Remote Sensing Magazie.

Project Description

Over the past several months, in synergy with my ESA point of contact, my PhD Supervisors, and other collaboratos, we have pioneered a series of quantum-enhanced methodologies aimed at transforming how Earth Observation (EO) data are processed, analyzed, and exploited.

Quantum Machine Learning (QML) is an emerging field in quantum computing research that has the potential to overcome several computational challenges in EO, such as developing accurate and efficient learning models. Our first deliverable was a detailed review article that mapped out the state of QML in the context of EO to present a detailed overview of the current state and trends in the application of QML to EO, examining its potential to improve data processing and analysis. We explored significant developments, such as in satellite image classification and segmentation, object detection and variable estimation. We also discussed the challenges, limitations, and future directions for integrating QML into EO applications, positioning it as a vital tool to advance our understanding of the dynamic systems of the Earth in the era of big data.

Building on the review’s insights, we developed Quanv4EO (Fig. 1), the first systematic quanvolutional preprocessing framework for multidimensional EO imagery. At its core is a frozen quantum kernel, which replaces a classic convolutional layer.

Fig. 1 Schematic of the proposed QuanvNN for EO data classification. On the left site the two quanvolutional layers and on the right side the three classical AI methods.

Quanv4EO matches or exceeds classical accuracy while using fewer trainable parameters.

This framework was also applied to tackle a critical environmental task: turbidity estimation in coastal waters. We fused quanvolutional preprocessing with a classical regression backbone, then trained on a novel dataset that pairs simulated Φ-Sat-2 spectral bands with Copernicus Marine Service turbidity measurements (Fig. 2).

Fig. 2 A graphical representation of the dataset processing pipeline and quantum deep learning architecture. The figure highlights the integration of the ESA OrbitalAI Challenge Simulator for generating ΦSat-2 spectral data, the use of High-Resolution Ocean Colour products from the Copernicus Marine Service as ground truth, and the application of a quanvolutional preprocessing step within a Fully Connected Neural Network framework, ultimately resulting in a predicted turbidity map.

The hybrid model reduced its parameter count by up to 98 % relative to a fully classical counterpart, yet delivers a 6.9 % boost in Pearson correlation and a 7.3 % reduction in RMSE.

Moreover, we evaluated the impact of integrating a quanvolutional preprocessing stage into an Attention U-Net architecture, focusing on Sentinel-1 SAR over Tunis, in order to perform an efficient building segmentation task. Preliminary experiments show that the hybrid network matches the segmentation accuracy of the baseline Attention U-Net while reducing trainable parameters by over 90 %.

Recognizing that fixed quantum kernels leave optimization potential untapped, we introduced a trainable quanvolutional layer (Fig. 3).

Fig. 3 Hybrid Quantum model for Binary Classification.

In this variant, the quantum parameters are updated via backpropagation alongside the classical network weights. On a binary EuroSAT classification task, the trainable quanvolution achieves even higher accuracy than both the frozen quantum and purely classical models, all while reducing trainable weights from ~96 k (frozen) to ~5 k (–94 %).

In these months we have also explored Generative models, since they are a pivotal for data augmentation, cloud removal, and inpainting in EO. Our Quantum Latent Diffusion Model (QLDM) embeds variational-quantum circuits into a 10-dimensional latent denoising process, replacing deep U-Net decoders with three quantum modules (Fig. 4).

Fig. 4 Architecture of the proposed quantum latent diffusion model

Among tested ansatzes (Rx-Rz-Rx, universal, matchgate), the universal circuit offers the best balance, reducing Fréchet Inception Distance by 21.5 % and Kernel Inception Distance by 29.9 % versus leading classical latent diffusion baselines, while preserving diversity (IS = 1.315).

Finally, we conducted a review on Quantum Graph Neural Networks, which represent a novel fusion of quantum computing and Graph Neural Networks, aimed at overcoming the computational and scalability challenges inherent in classical models that are powerful tools for analyzing data with complex relational structures but suffer from limitations such as high computational complexity and over-smoothing in large-scale applications. Quantum computing, leveraging principles like superposition and entanglement, offers a pathway to enhanced computational capabilities. We critically reviewed the state-of-the-art in Quantum Graph Neural Networks, exploring various architectures. We discussed their applications across diverse fields such as high-energy physics, molecular chemistry, finance and earth sciences, highlighting the potential for quantum advantage. Additionally, we addressed the significant challenges faced by Quantum Graph Neural Networks, including noise, decoherence, and scalability issues, proposing potential strategies to mitigate these problems. This comprehensive review aims to provide a foundational understanding of Quantum Graph Neural Networks, fostering further research and development in this promising interdisciplinary field.