Multitask Learning for Earth Observation Data Classification with Hybrid Quantum Network

TL;DR

The paper tackles EO data classification under Big Data constraints by introducing a multitask-based hybrid quantum neural network (MLTQNN) that uses an auxiliary image reconstruction task to enable efficient quantum encoding and a Location Weight Module to enhance quantum-convolution feature extraction. The architecture combines a quantum-feature extractor with a classical classifier and is trained with a multi-objective loss balancing $L_{ce}$ and $\alpha L_{mse}$, enabling robust learning even with limited data. Across four EO benchmarks, MLTQNN achieves higher test accuracy with fewer parameters than competitive baselines, and extensive ablations and validity analyses suggest the quantum component yields more meaningful, generalizable features. These findings highlight the potential of QML in EO data analysis, particularly for improving generalization and data efficiency in remote-sensing classification tasks.

Abstract

Quantum machine learning (QML) has gained increasing attention as a potential solution to address the challenges of computation requirements in the future. Earth observation (EO) has entered the era of Big Data, and the computational demands for effectively analyzing large EO data with complex deep learning models have become a bottleneck. Motivated by this, we aim to leverage quantum computing for EO data classification and explore its advantages despite the current limitations of quantum devices. This paper presents a hybrid model that incorporates multitask learning to assist efficient data encoding and employs a location weight module with quantum convolution operations to extract valid features for classification. The validity of our proposed model was evaluated using multiple EO benchmarks. Additionally, we experimentally explored the generalizability of our model and investigated the factors contributing to its advantage, highlighting the potential of QML in EO data analysis.

Figures (7)

• Figure 1: Overall architecture of the proposed model: image classification as the main task and image reconstruction as the auxiliary task
• Figure 2: The Quantum Circuit Example for Encoding: 1) $p_{x,y}^{i}$ indicates $i^{th}$ feature value of the superpixel $p_{x,y}$ 2) color indicates the quantum register: white qubits for $q_{l}$ and black qubits for $q_{v}$ 3) H denotes the Hadamard gate, while RX, RY, and RZ indicate rotation gates.
• Figure 3: The Quantum Circuit Example for Feature Extraction: 1) color indicates the quantum register: white qubits for $q_{l}$, black qubits for $q_{v}$, green qubits for $q_{k}$ and yellow qubits for $q_{f}$ 2) U denotes the gate unit consisting of RX,RY and RZ gates 3) dot markers indicate the controlled state: white dots for $\ket{0}$ and black dots for $\ket{1}$ 4) LWM stands for Location Weight Module, newly introduced to enhance feature extraction
• Figure 4: Illustration of the quantum convolutional operation using a $2\times2$-sized kernel with weights from $W_{0}$ to $W_{3}$: the quantum states $\ket{x_{2}x_{1}y_{2}y_{1}}$ in $q_{l}$ represent the locations of a $4\times4$-sized image, and the state $\ket{x_{1}y_{1}}$ can be used to identify the pixels transformed with the same weights. The visualization highlights $W_3$ as an example, where the state $\ket{11}$ selects its corresponding pixels for transformation
• Figure 5: model structure for the reconstruction task