Hybrid quantum-classical U-net for methane leaks detection

Overview

Methane is a short-lived but very strong greenhouse gas, and catching point-source leaks from oil and gas sites early matters. Deep learning (mostly U-Net) already segments methane plumes in satellite images, but quantum machine learning has barely touched image segmentation, and earlier quantum attempts used a quantum layer only as preprocessing in front of an unchanged network.

QCU-Net takes a different route: quantum convolution layers are placed directly inside several stages of the U-Net, and the whole model is trained end to end. It was developed in the Womanium Global Program 2024 (Quantum and AI Projects for Climate Change track) and is the subject of an accepted paper.

Key components and technical implementation

• Quanv layers: a quantum convolution slides a window over the feature map. Each pixel is angle-encoded onto one qubit, a trainable two-qubit circuit (tree-tensor-network style, RX or RY rotations plus CNOT) correlates neighbouring pixels, and the Pauli-Z expectation of each qubit becomes one output channel. Quanv 2×2 uses 4 qubits and Quanv 3×3 uses 9.
• QC-Input block: Quanv 2×2, then a classical 3×3 convolution to 9 channels, batch normalization and ReLU.
• QC-Down block: Quanv 3×3 in the first encoder stage, then a 64-channel classical convolution, batch normalization, ReLU and max-pooling. The remaining encoder stages and the whole decoder stay classical, because simulating quantum layers at every depth is too expensive.
• Variants compared: Archi1 (QC-Input only), Archi2 (QC-Input + QC-Down) and a classical U-Net with matching channel counts.
• Data: Sentinel-2 imagery (13 bands) around 21 oil and gas stations (7 in Algeria, 5 in Turkmenistan, 9 in the US) from a public dataset, cropped to 90×90 tiles with flips and rotations. Small (250 images) and medium (400 images) subsets were used.
• Training: PyTorch with the PennyLane default.qubit simulator, Adam, binary cross-entropy loss, and a grid search over learning rate, ansatz (RX or RY) and band subset (all 13 bands or bands 12–13, the methane-sensitive SWIR ones).

Key research findings and achievements

• QCU-Net variants beat the classical baseline on validation loss at both scales, and their validation curves followed the training curves smoothly, while the classical model’s validation loss oscillated.
• The gain is modest and shrinks with more data (gap of 0.034 on the small set, 0.0048 on the medium set). Adding the QC-Down block mattered more as the dataset grew.
• Results come from a noiseless simulator, so they say nothing yet about real quantum hardware. Evaluation on the large subset (1,200 images) is future work.
• The quantum bottleneck variant was left out of the results after a post-acceptance review found an implementation issue. The corrected version is planned future work.
• Achievement: to our knowledge, the first architecture to embed QCNN layers in multiple stages of a U-Net for end-to-end segmentation.