Quantum ReLU activation function for remote sensing classification and noise robustness analysis

Image classification systems based on convolutional neural networks (CNNs) are frequently subject to noise interruption, and the robustness of CNNs is influenced by the selection of activation functions (AFs). In this paper, we improved the quantum AFs QReLU and m-QReLU, making the parameter value range $ 0.00001 \le \alpha \le 0.01 $. We investigated the classification performance and noise robustness of quantum AFs using the EfficientNetB0, DenseNet121, and 2D-CNN models on remote sensing datasets. The experimental results showed that when the quantum AFs parameter is 0.0001, the classification accuracy of the model is higher than that of the ReLU AF. Under UC Merced and WHU RS19 datasets, the classification accuracy of quantum AF increased by 0.52% and 1.65% in the EfficientNetB0 model, and increased by 0.8% and 0.83% in the DenseNet121 model. Under the India Pines dataset, the classification accuracy of quantum AF increased by 1.71%. Furthermore, we individually introduced noise to the remote sensing datasets to investigate the robustness of quantum AFs. The results showed that when the quantum AFs parameter is 0.0001, the model has strong noise robustness.