Satellite image classification is a crucial task in remote sensing for applications such as land cover mapping, environmental monitoring, and resource management. While classical machine learning models such as SVMs achieve strong accuracy, they are often sensitive to noise, parameter-heavy, and less eficient when training data is scarce. In this work, we propose a hybrid quantum-classical pipeline for binary land cover classification on the EuroSAT dataset, combining a convolutional autoencoder for feature extraction with a Variational Quantum Classifier (VQC) inspired by quantum convolutional neural networks. The proposed QCNN-VQC achieves 90.0% accuracy, but more importantly outperforms baselines on F1-score (0.9428 after threshold tuning), AUROC (0.9819), and AUPRC (0.9812), while requiring fewer parameters. The model also demonstrates superior robustness to label noise and attains exceptionally high recall for the SeaLake class (97.8%), ensuring reliable detection of water bodies. These results highlight the promise of quantum machine learning for practical remote sensing tasks, especially in noisy or low-data regimes.
Satellite image classification plays a critical role in Earth observation, supporting applications ranging from agricultural monitoring and land-use mapping to climate change studies and disaster management. Traditional machine learning models, particularly Support Vector Machines (SVMs) and Convolutional Neural Networks (CNNs), have achieved notable success in remote sensing. However, these methods often require large volumes of annotated data, sufer performance degradation under label noise, and involve significant computational overhead.
Quantum Machine Learning (QML) has emerged as a promising alternative, ofering new ways to encode and process data through quantum states, entanglement, and variational circuits. By leveraging quantum resources, QML can potentially deliver greater expressive power with fewer parameters, improved generalization, and robustness in data-scarce settings. Recent advances in hybrid frameworks—where classical feature extractors are combined with quantum classifiers—have shown encouraging results for vision and geospatial tasks.
In this paper, we propose a hybrid quantum–classical framework for land cover classification using the EuroSAT dataset. A convolutional autoencoder compresses satellite images into a compact latent representation, which is then encoded into quantum states and classified using a QCNN-inspired Variational Quantum Circuit. While raw accuracy of 90.0% is slightly below SVM baselines, the quantum model demonstrates superior F1 (0.9428), AUROC (0.9819), AUPRC (0.9812), and robustness to label noise, alongside remarkable recall for the SeaLake class (97.8%). These results underscore the practical advantages of QML in remote sensing scenarios where false negatives, noise resilience, and parameter eficiency are critical.
Land cover mapping from satellite imagery is traditionally tackled with supervised learning. Given multi-spectral image inputs, models predict the land cover class for each pixel or image patch. Modern deep learning approaches, like CNNs, have become the state-of-the-art for such tasks, achieving high accuracy on benchmark datasets. For instance, Helber et al. created the EuroSAT dataset and demonstrated a deep CNN could classify ten land-use classes with over 98% accuracy on the test set. This success underscores the power of classical ML, but also sets a high bar for any novel approach like QML to be competitive.
Quantum Machine Learning (QML) leverages quantum computing for machine learning tasks. Notable QML models include parameterized quantum circuits (PQCs)—also called variational quantum circuits or quantum neural networks—and quantum kernel methods. In PQCs, data is encoded into quantum states, then a circuit with trainable gate parameters is executed, and measurements yield the model output. The parameters are optimized (with a classical optimizer) to minimize a loss, analogous to training a neural network. Quantum kernel methods, on the other hand, map data to a quantum state via a fixed feature-map circuit and use the quantum device to compute inner products (kernel values) between data points in this quantum feature space. Havlíček et al. first demonstrated that a quantum feature map can enable a support vector machine (SVM) classifier to separate classes that were hard to separate classically. These approaches are particularly relevant to image classification, as images have high-dimensional features that might benefit from quantum representations.
Given the limitations of today’s quantum hardware (noisy qubits and small quantum memory), hybrid models that combine classical preprocessing with quantum models have gained traction. A common pattern is to reduce the input data dimensionality using classical techniques (e.g., downsampling, principal component analysis, or neural network autoencoders) before feeding data to a quantum model. This ensures the data fits within the few qubits available and mitigates quantum noise by simplifying the task. There is increasing recognition that quantum components should complement classical ones—leveraging classical strengths in data preprocessing and using quantum circuits for specific parts of the pipeline.
Sebastianelli et al. [
Otgonbaatar and Datcu [
Delilbasic et al. and Zardini et al. [
Zöllner et al. [
Rodríguez-Grasa et al. [
Overall, prior work suggests that QML models can learn to classify remote sensing images with accuracy comparable to conventional methods, at least for small-scale or simplified tasks. Key lessons include the eficacy of hybrid architectures, the necessity of aggressive dimensionality reduction, and the observation that current quantum models do not yet vastly surpass classical accuracy but can sometimes match or slightly exceed it in controlled experiments. These studies form the foundation for our exploration of quantum machine learning for land-cover mapping.
The proposed pipeline integrates classical dimensionality reduction with a quantum variational classifier, as shown in Fig. 1. The methodology consists of four key stages: (i) data preparation and feature extraction via a convolutional autoencoder (CAE), (ii) quantum data encoding, (iii) variational quantum circuit (VQC) design, and (iv) measurement and classification.
Input EuroSAT images of size 64 × 64 × 3 are preprocessed by normalization and reshaping. To address the curse of dimensionality, a convolutional autoencoder (CAE) is employed to learn compact latent representations.
Let x ∈ R denote the flattened image vector. The encoder network ℰ(·) , parameterized by weights , maps the input to a lower-dimensional latent feature vector x̃︀ ∈ R :
Here, corresponds to the number of qubits in the quantum classifier. The decoder reconstructs the input, but only the encoder’s latent features x̃︀ are forwarded to the quantum stage.
The latent vector x̃︀ = (̃︀1, ̃︀2, . . . , ̃︀ ) is mapped to a quantum Hilbert space using angle encoding. Each feature is encoded as a rotation around the -axis: |( x̃︀, Θ)⟩ = ︃( )︃ ∏︁ (Θ ) enc(x̃︀)|0⟩⊗ =1 where Θ = {Θ of qubits.
A measurement is performed on the first qubit using the Pauli- operator. The model prediction is given by the expectation value: 1, . . . , Θ } are the trainable circuit parameters, and |0⟩⊗ is the all-zero initial state x̃︀ = ℰ(x),
≪ enc(x̃︀) = ∏︁ ()
̃︀ =1 where () = exp(− /2)
is the rotation gate around the Pauli- axis applied on the -th qubit.
On top of the encoded state, a variational quantum circuit is constructed using Strongly Entangling Layers (SEL). Each layer consists of parameterized single-qubit rotations followed by entangling gates across qubits.
Let (Θ ) denote the unitary operation for the -th layer with learnable parameters Θ . For layers, the variational state is given by:
To map this to a probability, the output is rescaled to [
x̃︀, Θ)⟩ ( = 1 | x̃︀, Θ) = The final class label is assigned by thresholding at 0.5. (1) (2) (3) (4) (5)
The variational parameters Θ are optimized via gradient-based methods (Adam optimizer) by minimizing the binary cross-entropy (BCE) loss:
=1 where are ground-truth labels and are predicted probabilities. Early stopping based on validation loss is applied to prevent overfitting.
ℒ(Θ) = − log() + (1 − ) log(1 − )
All experiments were conducted in Google Colab using TensorFlow 2.15, scikit-learn 1.5, and PennyLane 0.36. We used the EuroSAT dataset and selected two representative land-cover categories — Annual Crop and Sea Lake. Images were resized to 64×64 RGB, and a convolutional autoencoder (CAE) compressed them into latent features of dimension K=8. These latent vectors were then used for classification with Logistic Regression, SVM (RBF), and the proposed QCNN-inspired Variational Quantum Classifier (VQC). A 70/15/15 stratified split was used for training, validation, and testing.
The proposed QCNN–VQC achieved 90.0% accuracy, which is close to the performance of the SVM baseline (92.8%) and superior to Logistic Regression (87.2%). However, when evaluating with more robust metrics such as F1-score (after threshold tuning), AUROC, and AUPRC, the quantum model clearly outperforms classical baselines. Table 1 summarizes performance. The VQC matches or exceeds strong baselines on F1, AUROC, and AUPRC, and achieves high recall for SeaLake, which is valuable for environmental monitoring. Table 2 shows that the quantum model achieves a high recall of 97.8% for the SeaLake class, with only four false negatives. This indicates that nearly all water-body instances are correctly identified, which is particularly important for environmental monitoring applications.
Beyond raw accuracy, the QCNN–VQC shows state-of-the-art AUROC (0.9819) and AUPRC (0.9812), both surpassing SVM. These metrics confirm that the quantum model produces well-calibrated decision boundaries and superior precision–recall trade-ofs, particularly valuable for imbalanced or noisy data. In summary, although the SVM achieves slightly higher raw accuracy, the proposed quantum model delivers superior F1, AUROC, AUPRC, and robustness, making it the strongest overall choice in practical scenarios where false negatives and noise resilience are critical. Figure 3 shows autoencoder reconstructions, demonstrating that salient features are retained in the compressed latent space. Although minor blurring is visible, the reconstructed images preserve essential class information, validating the CAE as an efective dimensionality reduction stage. To assess separability, we applied t-SNE to the latent vectors. Figure 4 shows clear clustering of the two classes, supporting the suitability of the learned features for quantum encoding. The quantum model used only 144 parameters (6 layers × 8 qubits × 3), compared to 181 for a parameter-matched MLP.
Under 20% label noise, QCNN–VQC accuracy dropped by only 0.6%, compared to 3.6% for MLP.
This indicates superior robustness of the quantum classifier to noisy labels. Table 4 shows calibration results. Both models produce reasonably calibrated probabilities, with the MLP slightly outperforming QCNN–VQC.
Figure 5 plots accuracy vs. training samples per class. The QCNN–VQC performs well with as few as 100–200 samples, showing potential in data-scarce scenarios. At higher data volumes, MLP maintains an advantage.
This paper presented a hybrid quantum–classical pipeline for land cover mapping using the EuroSAT dataset. The proposed QCNN-inspired Variational Quantum Classifier achieved 90% accuracy and a best-threshold F1 of 0.9428, while delivering superior AUROC (0.9819), AUPRC (0.9812), and high recall for the SeaLake class (97.8%). These results demonstrate that hybrid quantum machine learning approaches can be not only competitive with strong classical models such as SVM, but can also surpass them under critical evaluation criteria such as noise robustness, parameter eficiency, and recall of minority-critical classes. Future work will extend this framework to multi-class land cover classification using the full EuroSAT dataset, incorporate multispectral bands beyond RGB, and explore larger remote sensing datasets such as RESISC45. Furthermore, as quantum hardware matures, deeper circuits with more qubits may unlock greater accuracy and scalability, reinforcing the promise of quantum-enhanced Earth observation systems.
The authors have not used Generative AI tools.
We thank colleagues and reviewers for constructive feedback.