Correcting Noisy Labels During the Classification of Hyperspectral Images

Classification is the basis of many remote sensing applications, however noise labeling largely affects the effectiveness of classifiers. To deal with this problem, this paper proposes a deep framework called Discerning Coteaching with Label Correction and Dropout(DCLCD), which can autonomously identify and correct noisy labels without additional prior knowledge. In detail, DCLCD uses a voting mechanism to ensure labels are accurately corrected after separating noisy labels. Furthermore, a dropout-like mechanism is included in both sample selection and label correction to enhance the robustness of the model. We validate the method on Salinas Valley and Pavia University, and the results reveal that DCLCD can reduce the proportion of noisy labels and improve classification accuracy while providing greater stability.