Practical insights on the effect of different encodings, ansätze and measurements in quantum and hybrid convolutional neural networks

TL;DR

This work systematically examines how three core PQC design elements—data encoding, variational ansätze, and measurement—affect quantum and hybrid CNN-like models applied to EuroSAT satellite imagery. By evaluating HQNN-Parallel, HQNN-Quanv, QCNN, and SEQNN across a comprehensive set of encoding-ansatz-measurement triplets, the authors isolate the relative impact of each block under fixed training conditions and against classical baselines. The results show a clear hierarchy: encoding choices dominate performance for hybrid models, while entanglement substantially influences HQNN-Quanv; pure amplitude-encoded models exhibit sensitivity to measurement and input ordering, achieving strong parameter efficiency. These findings provide practical guidance for designing quantum and hybrid image classifiers in remote sensing, while acknowledging the limitations of noiseless simulations and dataset scope.

Abstract

This study investigates the design choices of parameterized quantum circuits (PQCs) within quantum and hybrid convolutional neural network (HQNN and QCNN) architectures, applied to the task of satellite image classification using the EuroSAT dataset. We systematically evaluate the performance implications of data encoding techniques, variational ansätze, and measurement in approx. 500 distinct model configurations. Our analysis reveals a clear hierarchy of influence on model performance. For hybrid architectures, which were benchmarked against their direct classical equivalents (e.g. the same architecture with the PQCs removed), the data encoding strategy is the dominant factor, with validation accuracy varying over 30% for distinct embeddings. In contrast, the selection of variational ansätze and measurement basis had a comparatively marginal effect, with validation accuracy variations remaining below 5%. For purely quantum models, restricted to amplitude encoding, performance was most dependent on the measurement protocol and the data-to-amplitude mapping. The measurement strategy varied the validation accuracy by up to 30% and the encoding mapping by around 8 percentage points.

Figures (22)

• Figure 1: Samples from the EuroSAT dataset introduced in Helber2019. Each sample is a $64\times64\times3$ RGB image obtained from the Sentinel-2 satellite.
• Figure 2: Prototypical quantum circuit scheme.
• Figure 6: Depiction of the quantum linear layer for processing vector $\mathbf{x}$ with two circuits. The circuits independently process their corresponding vector chunk and can be run in parallel.
• Figure 7: Scheme of every model used. The three purely quantum models, \ref{['qcnnfc']}, \ref{['twokernel']} and \ref{['fcmodel']}, are particular cases of the generic PQC structure in Fig. \ref{['quantumlayer']}. Quantum circuits were created using draw.io, an open-source library drawio.
• Figure 8: Validation accuracy of all configurations for a fixed encoding.