Robustness of convolutional neural network models in hyperspectral noisy datasets with loss functions

Abstract The presence of noise in images affects the classification performance of convolutional neural networks (CNNs). The loss function plays an important role in the noise robustness of CNN models. Loss sensitivity represents the loss function robustness for noisy data. In this study, the robustness of CNN models was investigated by using the cross-entropy, pseudo-Huber, and correntropy loss functions on noisy data. The experiments were performed using the Xception architecture for Pavia University and Salinas Scene datasets. Some common noises in hyperspectral images (HSIs), such as Gaussian, stripe, and salt-and-pepper noises, were applied to the test data, and the results of classification with different loss functions were compared. To reduce the training time and prevent overfitting, a HSI pixel-to-image sampling method was proposed. According to the results, the correntropy loss function was more robust to noises, while the cross-entropy loss function was more accurate for noiseless data as compared to the other loss functions.