Quantum Machine Learning for Earth Observation Images

In this work, we present our first study on Quantum Machine Learning (QML) applied to the Earth Observation (EO) domain, which has consistently leveraged state-of-the-art advances in Machine Learning for imagery, including recent researches on QML based techniques. Based on Ref. [1], we suggest an alternative approach for Quantum Convolutional Neural Networks (QCNN) and test its performance as image classifiers with different single-qubit and two-qubit gates for the EuroSAT dataset. As the original image of size 64x64 is too large for the current quantum hardware, we reduce its dimension through classical feature extractions methods, PCA, and convolutional autoencoder. We start with a binary classification of real EuroSAT features and fake data which are sampled from a uniform distribution. Then, we perform a set of binary classification between different couples of EuroSAT classes to investigate its capability for more detailed classification. Finally, we use the QCNN architecture as a 4-class classifier for MNIST and EuroSAT datasets and compare their results. Although improvements are still required in some areas, especially for the EuroSAT dataset, our preliminary work shows the potentialities of using QCNN as a quantum classifier in the context of computer vision and in particular Earth observation. Our ultimate goal is to extend this QCNN to generative models, including Generative Adversarial Networks (GAN), by replacing the generative part, and possibly the discriminative part, with quantum circuits.

[1] T. Hur, L. Kim, and D. K. Park. Quantum convolutional neural network for classical data classification, 2021.