Performance Evaluation of a One-Dimensional Quantum Convolutional Neural Network for Remote Sensing

Quantum Machine Learning (QML) continues to gain attention as a possible solution to the data-scarcity challenges that limit classical machine learning in hyperspectral remote sensing. With the emergence of accessible quantum processors, evaluating the practical performance of quantum neural networks has become increasingly relevant. This study investigates the use of a classical convolutional neural network (CNN) and a one-dimensional quantum convolutional neural network (QCNN) for classifying Holm Oak trees using PRISMA hyperspectral imagery, with particular focus on extremely small training sets. Using a dataset of 20 labeled training samples, the QCNN was implemented on both a quantum simulator and IBM’s superconducting quantum hardware. Performance was assessed in terms of classification accuracy, number of qubits/feature dimensions, and training and prediction time. Results showed that: (i) with 20 training samples, the QCNN achieved similar accuracy to the CNN from 4 to 14 qubits/features; (ii) QCNN training time with 20 samples was approximately 50 times higher than the GPU-accelerated CNN due to the overhead of optimizing variational quantum circuits; and (iii) prediction time for 400 pixels was substantially slower for the QCNN across all qubit counts. Overall, despite achieving comparable accuracy in most configurations, the QCNN demonstrated poor computational performance compared to the classical CNN, which remained significantly faster for both training and inference.