HQ-UNet: A Hybrid Quantum-Classical U-Net with a Quantum Bottleneck for Remote Sensing Image Segmentation
Quantum circuit boosts segmentation accuracy to 94.76% with fewer parameters.
A new paper from researchers including Md Aminur Hossain introduces HQ-UNet, a hybrid quantum-classical architecture that integrates a compact parameterized quantum circuit (PQC) into the bottleneck of a classical U-Net for remote sensing image segmentation. The design replaces the typical pooling-based encoding with a non-pooling quantum convolutional module that enriches highly compressed features before decoding. This approach keeps the quantum component shallow and parameter-efficient, addressing the challenge of applying quantum machine learning to high-dimensional satellite images on near-term quantum hardware.
On the ISPRS Vaihingen dataset, HQ-UNet achieves a mean intersection-over-union (IoU) of 0.8050 and overall accuracy of 94.76%, surpassing the classical U-Net baseline. The results highlight that even a small quantum bottleneck can enhance feature representation for dense prediction tasks. While still experimental, the work points to a practical path for leveraging noisy intermediate-scale quantum (NISQ) devices in Earth observation, potentially enabling more accurate land-use classification, disaster monitoring, and urban planning with fewer trainable parameters.
- Achieves 94.76% overall accuracy and 0.8050 mean IoU on the ISPRS remote sensing dataset, outperforming classical U-Net.
- Uses a compact parameterized quantum circuit at the U-Net bottleneck with a non-pooling quantum convolutional module.
- Keeps quantum component shallow and parameter-efficient, making it viable for near-term NISQ hardware constraints.
Why It Matters
Demonstrates practical near-term quantum advantage for high-dimensional satellite imagery analysis, enabling more accurate Earth observation with fewer parameters.