🏛️ Company / Organization | KP Labs (PL) – Prime Contractor, IT4Innovations (CZ), Jagiellonian University (PL) |
📆 Contract Duration with ESA Φ-lab | May 2025 - November 2026 |
🌍 Project Title | Gaussian Boson Sampling for Earth Observation (GBS4EO) |
🌐 Project Website |
Project Description
Abstract
Quantum computing (QC) is advancing with various hardware technologies at different stages. Some aim for universal quantum computers, while others focus on specialized applications. Technologies like quantum annealing and photonic quantum computers offer efficient solutions, with photonic quantum computers being promising due to their simplicity and robustness. Remote Sensing (RS) and Earth Observation (EO) technologies are advancing due to technological innovation, market demand, and collaboration, offering opportunities across a range of EO downstream tasks. The increasing data availability drives the need for innovative RS data analysis technologies – machine learning (ML) has grown rapidly in this context. Quantum technologies also show promise for computational challenges but face significant technical hurdles.
The aim of the Gaussian Boson Sampling for Earth Observation (GBS4EO) project has been to bridge quantum ML and EO, focusing on real use cases such as hyperspectral band selection and pattern matching for methane leak detection. It is also applicable to other tasks involving the analysis of imbalanced, high-dimensional EO data.
Non-classical states of light span an infinite Hilbert space, offering significant potential for QC applications. Unlike a single qubit in a 2-dimensional Hilbert space, a single mode of light is described within an infinite Fock space. This, combined with light's long coherence times at room temperatures and photonic technologies, gives this QC approach unique advantages. Gaussian Boson Sampling (GBS) uses non-classical states of light and builds on the concept of Boson Sampling. Classical computers cannot simulate GBS outputs in polynomial time, but recent experiments with programmable optical circuits have demonstrated GBS's advantages. Consequently, GBS presents a validated method for surpassing classic capabilities, with tangible implementations. It is distinct amidst ongoing efforts to realize practical quantum computers. While GBS does not provide a foundation for universal QC, it excels in specific tasks (e.g., graph problems), which we have built upon in GBS4EO.
The primary objective of GBS4EO has been to build end-to-end solutions that benefit from the GBS for satellite data analysis. A key advantage should be the availability of GBS solutions, enabling immediate integration – a feat that is challenging to achieve with alternative QC approaches due to technical complexities. Our investigation has encompassed theoretical inquiries, numerical experiments, and experiments on a quantum computer. We have devised a compiler capable of translating the GBS problem onto the architecture of universal quantum computers. This endeavor necessitates mapping squeezed states and operations executed within linear optical circuits onto the language of qubits and universal quantum circuits. Such an approach has enabled us to evaluate GBS-based methodologies on universal quantum computers, including those developed by European initiatives. We have also leveraged a classical simulator to develop GBS-enhanced satellite data analysis – the promising outcomes will pave the way for eventual deployment on physical GBS photonic chips. Since the consortium members have access to the VLQ quantum computer, we have also used it to simulate GBS.
Detailed description
The GBS4EO project was implemented through a sequence of technical activities covering the preparation of Earth Observation datasets and validation procedures, the development of quantum algorithms and supporting software, and the experimental validation of the resulting solutions. The work was organized into three main technical areas: Datasets and Validation Procedures, Quantum Algorithms, and Experimental Validation. Together, these activities establish the complete workflow required to develop and assess Gaussian Boson Sampling-based approaches for selected Earth Observation applications.
- AI-ready datasets and validation procedures
The first part of the project established the data, validation, and classical machine-learning foundation required for the subsequent development and assessment of the quantum algorithms.
A major outcome of this work was the preparation of AI-ready datasets suitable for the experiments planned within GBS4EO. For methane detection, the dataset preparation included hyperspectral imagery accompanied by the corresponding ground-truth information required for validation. For hyperspectral band selection, datasets were prepared for several downstream Earth Observation tasks.
The datasets were accompanied by documented training and test splits, as well as validation procedures. This was important to ensure that the same data partitions, metrics, and evaluation principles could subsequently be applied to both classical and quantum-based approaches. Establishing a consistent experimental basis from the beginning of the project was necessary to support meaningful comparisons between different algorithms.
The project also defined the quality metrics to evaluate the considered Earth Observation tasks and established classical machine-learning benchmarks. Relevant classical algorithms were implemented and evaluated using the prepared datasets and the agreed validation procedures. These results provide the reference performance against which the Gaussian Boson Sampling-based approaches developed later in the project can be assessed.
The resulting work therefore delivered two main categories of output. The first consisted of AI-ready datasets for methane detection and hyperspectral band selection, together with the corresponding validation procedures. The second consisted of experimental results obtained using classical machine-learning algorithms across the selected tasks and datasets.
Practical examples were also prepared to support the reproducibility and usability of the project results. These examples demonstrate how the prepared datasets can be handled and how the classical machine-learning models and benchmark results can be reproduced.
This part of the project, therefore, provided both the experimental data foundation and the classical reference results required for the subsequent development and validation activities.
- Quantula repository
The project has produced a set of complementary software outcomes. Within this ecosystem, the Quantula repository provides the Earth Observation algorithmic layer for GBS4EO. The repository implements modular Python workflows for GBS-based hyperspectral feature selection and methane-plume segmentation, together with supporting tools for graph construction, sampler integration, experiment execution, validation, and result analysis. Quantula is designed to interoperate with companion project software, including the gate-based optical-circuit simulation route supported by GBS Compiler. This creates a software bridge between Earth Observation use cases and hardware-oriented representations of optical sampling circuits.
The codebase was designed around the central idea of the project: reformulating selected Earth Observation machine-learning tasks as graph problems whose relevant structures can be sampled with a Gaussian Boson Sampling primitive. In this framework, hyperspectral bands, image pixels, or local pixel-reference structures are represented as graph nodes, while edges encode similarity or affinity. GBS-based sampling is then used to identify informative dense substructures, which are converted into selected features, pixel-level evidence, or segmentation labels.
Quantula currently includes implementations for two main application tracks: hyperspectral feature selection and methane detection / semantic segmentation. It also includes data-handling utilities, experiment runners, validation scripts, result-aggregation tools, and documentation for reproducing the latest benchmark workflows. The repository supports a separation between the algorithmic layer and the sampler backend. In the current version, experiments can be run not only with classical GBS simulation backends but also via a gate-based optical-circuit simulation route developed in connection with the companion project repository, GBS Compiler.
- GBS Compiler component
The gbs-compiler repository is a companion software outcome of the project, focused on the gate-based representation and simulation of optical circuits relevant to Gaussian Boson Sampling. It complements Quantula by providing the lower-level compiler layer: Quantula implements Earth Observation workflows, while gbs-compiler translates continuous-variable optical operations into qubit-based circuit representations that can be simulated or exported in a gate-based quantum computing format.
The repository is based on a truncated Fock-space encoding of optical modes. In this approach, each optical mode is represented by a finite number of photon-number states, which are encoded into the computational basis of qubits. The current implementation supports photon-number cutoffs of 2 and 4, corresponding to one or two qubits per optical mode. Within this encoding, standard optical components such as phase shifts, displacements, squeezing operations, and beam splitters are decomposed into qubit gates.
A key role of gbs-compiler is to provide a practical compilation route from photonic Gaussian operations to gate-based circuits. The documented pipeline starts from Blackbird source descriptions of optical circuits, parses them into operation lists, decomposes each continuous-variable operation into qubit gates, constructs a PennyLane quantum script, and exports the resulting circuit as OpenQASM 2.0. This creates a concrete software bridge between photonic Gaussian Boson Sampling models and hardware-oriented circuit descriptions.
- GBS-based hyperspectral feature selection
The first algorithmic outcome is a GBS-based approach to hyperspectral feature selection. The method addresses a standard challenge in hyperspectral Earth Observation: many images contain a large number of spectral bands, but only a smaller subset may be needed for a given downstream task. Selecting informative bands can reduce computational cost, improve interpretability, and limit redundancy while retaining task-relevant information.
The project approach represents candidate hyperspectral bands as nodes of a similarity graph. Edges are created using a distance or affinity measure that combines statistical and spatial information about the bands. In the technical design, graph construction can use measures such as information variation and spatial local-variance descriptors, allowing the method to account not only for spectral redundancy but also for differences in image structure. Once the graph is constructed, the GBS-based subgraph routine is used to sample dense or otherwise informative band subsets.
The selected features can then be passed to classical downstream models for validation on classification, regression, unmixing, or segmentation tasks. This makes the algorithm useful both as a quantum-inspired dimensionality-reduction tool and as a bridge between near-term quantum sampling primitives and conventional Earth Observation machine-learning pipelines.
A practical outcome of this work is an implementation pathway in Quantula that separates the feature-selection logic from the sampler backend.
- Spatial-spectral methane detection and semantic segmentation
The second algorithmic outcome is a GBS-based methane-detection pipeline for pixel-level segmentation of airborne hyperspectral scenes.
The algorithm does not train a conventional end-to-end segmentation model. Instead, it performs local evidence accumulation. For each test crop, small local groups of query pixels are sampled according to spatial neighborhoods. These query pixels are then combined with external positive and negative reference anchors drawn from labeled training scenes. The resulting local graph connects query pixels using spatial--spectral similarity analysis, while the reference anchors provide methane/background evidence. A fast GBS subgraph-selection backend samples informative local graph structures, and the resulting evidence is accumulated into hard methane labels, soft methane scores, and evidence-coverage maps.
The latest HPC experiments demonstrate the pipeline's technical viability on real preprocessed STARCOP data, using eight selected AVIRIS bands together with the MAG1C methane-prior product. The results show that the method can produce meaningful methane-plume evidence maps and useful diagnostic outputs, including soft methane scores and evidence-coverage maps.
The figure illustrates two favorable cases in which the method recovers the main plume morphology and produces a meaningful soft methane-evidence map.
- Experimental validation
Eventually, it is all about combining the outputs of the preceding activities and addressing their integration and experimental assessment.
The first stage of this work focused on the preparation of a robust and reproducible experimental pipeline for quantum machine-learning experiments. The experimental environment provides the framework required to conduct systematic experiments using high-performance computing resources and a VLQ quantum computer (https://www.it4i.cz/en/infrastructure/vlq-quantum-computer). It also supports consistent management of the software, experimental configurations, models, results, and other relevant artifacts generated during the validation process.
Particular attention was given to reproducibility. The experimental pipeline was therefore developed to support the systematic execution and tracking of experiments and to ensure that the results obtained can be analyzed and reproduced consistently.
The resulting experimental pipeline has been documented in detail, including its architecture, main assumptions, and data flows. Completion of this activity provides the technical environment required to conduct the final validation campaign using the datasets, classical benchmarks, and quantum algorithms developed in the earlier stages of the project.
The remaining activity is the experimental validation of the developed quantum machine-learning algorithms. This work is currently in progress.
Its purpose is to execute the planned experiments using the integrated GBS4EO environment and to assess the developed algorithms across the selected Earth Observation datasets and downstream tasks. The validation includes comparison of the quantum approaches with the classical machine-learning baselines established earlier in the project.
The final validation report will consolidate these results and provide a detailed assessment of the behavior of the investigated quantum machine-learning approaches. In particular, it will address any observed potential advantages of the Gaussian Boson Sampling-based methods, while also identifying limitations and open technical issues that would need to be addressed to support further adoption of quantum machine learning in Earth Observation.
As this final experimental activity is still ongoing, conclusions concerning the comparative performance of the developed algorithms and any potential quantum advantage will be shared upon the project completion.
Project Outcomes
Development outputs:
- Quantula repository ➜
- GBS Compiler repository ➜
https://github.com/kplabs-pl/Quantula
Scientific publications:
- Presentation given at 7th Quantum Technology Conference 2025 ➜