PaperlayerView exact revision on arXiv

Source-linked AI summary

On Circuit-based Hybrid Quantum Neural Networks for Remote Sensing Imagery Classification

Alessandro Sebastianelli, Daniela A. Zaidenberg, Dario Spiller, Bertrand Le Saux, Silvia Liberata Ullo

arXiv:2109.09484v2eess.IVcs.CVcs.ETquant-ph

TL;DR

As Earth Observation image volumes grow, the paper investigates whether hybrid quantum convolutional networks can classify land use and land cover imagery. It evaluates QCNNs on EuroSAT and reports better multiclass performance than classical counterparts, with entangled circuits performing best.

  • Problem

    The paper examines whether quantum computing can improve machine-learning performance for LULC classification in remote sensing, where image volumes are growing.

    Source: §I · pp. 1–2

  • Method

    The study evaluates circuit-based hybrid QCNNs, which add a quantum layer to a standard neural network, on the EuroSAT remote-sensing classification task.

    Source: §I · p. 2, §VII · p. 10

  • Results

    On EuroSAT, hybrid QCNNs outperformed their classical counterpart in multiclass classification, and entangled circuits achieved the best scores.

    Source: §I · p. 2, §VII · p. 12

  • Takeaways & Limitations

    The authors present the EuroSAT experiment and open-access code as a resource for geoscience and remote-sensing researchers investigating quantum computing for Earth Observation problems.

    Source: §VII · p. 12

Abstract

from arXiv · show

This article aims to investigate how circuit-based hybrid Quantum Convolutional Neural Networks (QCNNs) can be successfully employed as image classifiers in the context of remote sensing. The hybrid QCNNs enrich the classical architecture of CNNs by introducing a quantum layer within a standard neural network. The novel QCNN proposed in this work is applied to the Land Use and Land Cover (LULC) classification, chosen as an Earth Observation (EO) use case, and tested on the EuroSAT dataset used as reference benchmark. The results of the multiclass classification prove the effectiveness of the presented approach, by demonstrating that the QCNN performances are higher than the classical counterparts. Moreover, investigation of various quantum circuits shows that the ones exploiting quantum entanglement achieve the best classification scores. This study underlines the potentialities of applying quantum computing to an EO case study and provides the theoretical and experimental background for futures investigations.

I. INTRODUCTION

The paper tests circuit-based hybrid QCNNs for LULC classification in Earth Observation, extending earlier remote-sensing quantum work beyond binary tasks. It reports better multiclass results than classical counterparts, with entangled circuits achieving the best scores.

I. INTRODUCTION

  • EO image volumes exceed 150 terabytes per day, pressuring image-classification techniques to keep pace.The introduction frames quantum computing as one technology being explored for this processing demand.Source: §I · p. 1
  • Existing quantum remote-sensing work includes satellite quantum imaging and annealer applications to SAR phase unwrapping, feature selection, and hyperspectral classification,,.Source: §I · p. 2
  • The authors position their study amid few remote-sensing QML works, including prior binary classification and a deep-convolutional classifier using a 17-qubit circuit.Source: §I · p. 2
  • The study applies quantum computing to EuroSAT optical multispectral imagery, going beyond early proofs of concept on a few images,.The paper also extends earlier remote-sensing quantum classification work beyond binary tasks.Source: §I · p. 2
  • Hybrid QCNNs achieve better multiclass results than fully classical counterparts, and entangled gate-based circuits achieve the strongest reported scores.The comparison was conducted on the EuroSAT benchmark.Source: §I · p. 2
  • The contribution also includes structured coarse-to-fine prediction and classical and quantum models designed or adapted to the dataset.Source: §I · p. 2

II. LAND USE LAND COVER CLASSIFICATION OVERVIEW

The overview describes why LULC classification is difficult across changing acquisition conditions and surveys the approaches used to address it. The QML discussion presents possible benefits alongside unresolved questions about advantage and quantum-computing constraints.

II. LAND USE LAND COVER CLASSIFICATION OVERVIEW

  • Classical pixel- and object-based LULC methods are affected by cloud, fog, sensor changes, and environmental variation, complicating generic classification across regions and land classes.Source: §II · p. 2
  • Deep-learning approaches include patch-wise and pixel-wise classification, with CNNs, random fields, FCNs, and newer capsule, recurrent, graph, and transformer models.The overview associates patch-wise methods with local neighborhoods and pixel-wise methods with local appearance statistics.Source: §II · p. 3
  • The overview frames quantum computing as a possible response to LULC challenges involving growing data availability, large-scale processing, sensor variety, and very high resolution.Source: §II · p. 3, §III · p. 3

III. QUANTUM MACHINE LEARNING

  • QML is explored for potential runtime, learning-capacity, and learning-efficiency improvements, with proposed mechanisms including quantum tunneling and quantum linear algebra, sampling, and optimization.Source: §III · p. 3
  • There is no consensus on when QML outperforms classical models across general problem classes, and data quality and quantity can affect their relative performance.Source: §III · p. 3
  • In the NISQ era, limits on quantum operations constrain scaling, although the overview notes ongoing work toward more reliable computation.Source: §III · p. 4

IV. MATHEMATICAL BACKGROUND ON QC

The section introduces qubits, measurement, and quantum gates as foundations for quantum circuits. It explains how entanglement arises and why it matters to the circuits investigated in the paper.

IV. MATHEMATICAL BACKGROUND ON QC

  • A qubit can exist in a coherent superposition of |0⟩ and |1⟩, while measurement irreversibly yields information and destroys that superposition.The qubit state is described by probabilities associated with the two states.Source: §IV · p. 4
  • Entangled states cannot generally be factorized into their original qubits, and the paper identifies entanglement as a building block of its investigated circuits.The Hadamard and CNOT gates are combined to create a Bell state that introduces entanglement.Source: §IV · pp. 4–5
  • For the Bell state, measuring the first qubit as |0⟩ leaves the second qubit in |0⟩ as well.The paper describes this as a consequence of entanglement, even when the qubits are separated by a very large distance.Source: §IV · p. 4
  • Quantum gates are unitary operators that preserve inner products and can be visualized as rotations of the quantum state on the Bloch sphere.The section introduces Hadamard, rotation, and CNOT gates.Source: §IV · pp. 4–5

V. METHODOLOGY

The methodology section outlines the hybrid QCNN components and its optimization and inference, and notes that the authors implemented the code from scratch.

V. METHODOLOGY

  • The methodology covers quantum circuits as potential layers, their integration with the classical architecture, and hybrid QCNN optimization and inference.It introduces data embedding and an example interface between classical and quantum layers.Source: §V · p. 5
  • The authors state that they implemented all code from scratch and released it in a public repository.The circuits are described as standard quantum-computing units and fundamental units of IBM Qiskit.Source: §V · p. 5

A. Data Embedding

Data embedding maps classical activations into a quantum representation for the hybrid network. A parametrized circuit processes that representation and passes measurement statistics to the following hidden layer.

A. Data Embedding

  • A higher-dimensional quantum representation of classical data is needed to integrate a parametrized quantum layer into a classical network.The section describes preparing a quantum state for this purpose.Source: A. Data Embedding · p. 5
  • Data embedding applies a unitary operator to N |0⟩ quantum nodes to encode classical information in an N-qubit space.The unitary's parameters are determined by the preceding classical nodes, and the encoding represents their activation through the probability of measuring |1⟩.Source: A. Data Embedding · pp. 5–6
  • The section describes Hadamard and RZ-gate rotations as one way to encode classical feature values into qubits.The rotation angles correspond to the feature values of preceding inputs.Source: A. Data Embedding · p. 6
  • A parametrized quantum circuit uses rotation angles specified by a classical input vector, with outputs from the preceding neural-network layer serving as its inputs.The circuit's measurement statistics are collected as inputs for the following hidden layer.Source: A. Data Embedding · p. 6

B. Selected Quantum Circuits for Image Classification

The study compares three four-qubit circuits, distinguished by whether and how they introduce entanglement. Their gate arrangements vary how classical inputs are encoded into quantum activations.

B. Selected Quantum Circuits for Image Classification

  • The three selected circuits use four qubits, with the first omitting entanglement and the remaining two introducing it through different gate connections.The authors note that the selection is illustrative rather than an exhaustive comparison of possible quantum configurations.Source: B. Selected Quantum Circuits for Image… · p. 6
  • The no-entanglement circuit applies Hadamard and rotation gates to encode classical inputs as quantum activations.The rotation angles are derived from preceding classical nodes.Source: B. Selected Quantum Circuits for Image… · p. 6
  • The Bellman circuit entangles qubits with Hadamard and sequential CNOT gates before applying input-parameterized y-axis rotations.The circuit then reverses the CNOT application while retaining superposition, transforming the encoding in quantum feature space.Source: B. Selected Quantum Circuits for Image… · p. 7
  • The Real Amplitudes circuit applies Hadamard gates and input-derived rotations, then mutually entangles the qubits using CNOT gates.The rotations encode classical information as quantum information before entanglement.Source: B. Selected Quantum Circuits for Image… · p. 7

C. Hybrid Quantum Neural Network Classifier

The paper adapts a classical CNN with a quantum circuit layer for remote-sensing image classification, and compares it with classical counterparts on EuroSAT.

C. Hybrid Quantum Neural Network Classifier

  • The proposed hybrid QCNN combines classical machine learning and quantum layers, a design the paper identifies as preferred amid current quantum-technology limitations.The authors transfer and adapt this hybrid approach from computer vision to remote sensing.Source: C. Hybrid Quantum Neural Network… · p. 7
  • Its classical CNN derives from LeNet-5, with convolutional layers and input dimensions changed to fit the images and two fully connected layers around the quantum layer.These layers adapt the input to quantum capacity and map the quantum output to the dataset's class count.Source: C. Hybrid Quantum Neural Network… · pp. 7–8
  • The quantum layer uses one of the described circuits, and several circuit designs are tested to investigate their potential.The layer is intended to benefit from probabilistic quantum computing properties.Source: C. Hybrid Quantum Neural Network… · p. 8
  • Two classical CNN comparators replace the circuit with either a 16-node fully connected layer or a multilayer perceptron with 256, 64, 32, and 10 nodes.The alternatives provide classical counterparts for comparison with the proposed QCNN.Source: C. Hybrid Quantum Neural Network… · p. 8
  • Experiments cover three quantum circuits, two classical models, coarse and fine-grain quantum classifiers, and a segmentation-map experiment.The dataset is EuroSAT, with 13 Sentinel-2 spectral bands, 10 classes, and 27,000 labeled, georeferenced images; it is split 80/20 for training and validation.Source: C. Hybrid Quantum Neural Network… · p. 8
  • All models were designed and implemented from scratch, adapted to the dataset, and trained without pretrained weights.Hyperparameters and loss settings were selected for the problem requirements.Source: C. Hybrid Quantum Neural Network… · p. 8

D. Training and testing

The hybrid networks train through classical and quantum computation, with testing on the validation set. The paper specifies shared circuit initialization and training settings for reproducibility.

D. Training and testing

  • All qubits in the three circuit designs are initialized to |0⟩.The paper states that training and testing use the same settings where possible.Source: D. Training and testing · p. 8
  • Training ran on Google Colaboratory with a Tesla K80 GPU, a single-core Xeon CPU, 12.6 GB available RAM, and 320 GB available disk.Source: D. Training and testing · p. 8
  • Each QCNN trains for 50 epochs with Adam, a 0.0002 learning rate, and cross-entropy loss; the classical CNNs take about 100 epochs to converge under the same training approach.Source: D. Training and testing · p. 8
  • Training sends image features through the CNN and quantum circuit, compares the output with ground truth, then updates model weights using the error and gradient.The described process has feed-forward and backward streams.Source: D. Training and testing · p. 8
  • Testing applies the trained model to the validation set and reports confusion matrices, accuracy, precision, recall, and F1 scores.The training and testing procedures are summarized in Algorithms 1 and 2.Source: D. Training and testing · p. 8

A. EuroSAT dataset classification

The Real Amplitudes QCNN achieves the strongest reported overall classification result, and the authors report better performance for entangled circuits than for classical counterparts. The no-entanglement model fails to recover Highway, a difficult class.

A. EuroSAT dataset classification

  • The reported evaluation includes confusion matrices and classification reports for accuracy, precision, recall, and F1 score.Table I reports per-class F1 scores and overall accuracy for three quantum and two classical classifiers; Tables II and III report per-class precision and recall.Source: A. EuroSAT dataset classification · p. 9
  • 92% accuracy, a +10% gain over the second best approach, is achieved by the Real Amplitudes circuit; its confusion matrix has nearly perfect diagonal scores and surpasses the other models.The authors report that both entangled circuits also perform better than the two classical counterparts.Source: A. EuroSAT dataset classification · pp. 9–10
  • The no-entanglement model fails to recover Highway, a class that all analyzed classifiers find difficult.The authors contrast this circuit's poor effectiveness on more complex remote-sensing images with its successful use for digit-image classification.Source: A. EuroSAT dataset classification · p. 10

B. Coarse-to-fine structured land-cover classification

The coarse-to-fine classifier combines a three-way coarse stage with fine-grain classifiers and improves on the standard classifier while matching the reported state-of-the-art accuracy. On unseen Beirut imagery, its patch-wise maps were comparable with other state-of-the-art solutions, with a slight residential-area advantage reported.

B. Coarse-to-fine structured land-cover classification

  • Three fine-grain Real Amplitudes classifiers separate visually similar classes within vegetation, urban, and water-body macro-classes.The method uses a four-qubit circuit within the selected macro-classes to encode finer distinctions.Source: B. Coarse-to-fine structured… · p. 10
  • The coarse classifier reached 98% overall accuracy and 98% overall F1.It divides the data into three macro-classes before the matching fine-grain classifier produces the final classification.Source: B. Coarse-to-fine structured… · p. 10
  • Fine-grain classifiers reported overall accuracies and F1 scores from 94% to 99%, consistently exceeding standard-classifier results for individual classes.These classifiers cover vegetation, urban, and water-body clusters; the comparison uses constant complexity on a slightly reduced dataset.Source: B. Coarse-to-fine structured… · p. 10
  • 97% overall accuracy and 97% overall F1 were reported for the coarse-to-fine classifier, improving over the standard classifier by +3% and matching Helber et al.'s 98.57% accuracy.The proposed architecture has six layers versus ResNet-50's 50, and is described as an asset for frugal-resource computations.Source: B. Coarse-to-fine structured… · p. 10

C. Semantic segmentation by patch-wise classification

  • The trained fine-grain quantum classifier was applied to unseen Sentinel-2 images from OSCD using 64x64-pixel sliding windows with a 32-pixel step, producing patch-wise semantic maps.This reproduced experiment for comparison with state-of-the-art deep-learning approaches.Source: C. Semantic segmentation by patch-wise… · p. 10
  • On unseen Beirut imagery, the patch-wise quantum classifier distinguished urban, vegetation, and water zones and was comparable with state-of-the-art maps, with a slight advantage in retrieving residential areas.The comparison used Wide-ResNet and JEM models.Source: C. Semantic segmentation by patch-wise… · p. 10

VII. CONCLUSION

The conclusion reports successful multiclass classification on EuroSAT, with entangled circuits outperforming other circuits and the quantum layer outperforming its classical counterpart. It identifies greater quantum processing and improved data encoding as future directions.

VII. CONCLUSION

  • The hybrid QCNN was effective for multiclass identification and computing efficiency in remote-sensing image classification.The approach adds a quantum layer to a standard neural network.Source: §VII · p. 10
  • On EuroSAT, entangled circuits achieved better results by a significant margin, and the quantum layer outperformed its classical counterpart.The experiments evaluated multiclass classification of Earth-observation scenes.Source: §VII · p. 12
  • Current NISQ capacity requires a classical component for data embedding; future work targets more powerful encoding networks and a larger share of quantum processing.Quantum convolutions are proposed for future study to incorporate spatial information and invariance.Source: §VII · p. 12
  • Understanding probabilistic mechanisms in quantum layers is identified as key to designing better models, developing deep quantum learning, and eventual real-life applications.The paper makes its code and experiments openly available on GitHub.Source: §VII · p. 12
Loading 2109.09484v2…