Ensemble Label Correction for Hyperspectral Image Classification

The presence of noisy labels in the training sets poses a challenge to achieving high-precision classification of hyperspectral images (HSIs). To address this issue, this paper proposes an ensemble label correction (ELC) method for HSI classification that incorporates the concept of ensemble learning into deep neural networks. Our ELC model utilizes ensembles to gradually correct determined noisy labels during training epochs. At the end of training, most wrong labels will be corrected, and the model's robustness will also be increased. In our experiments, we use convolutional neural networks to implement our ELC model. Our experiments on two real HSI datasets illustrate that the proposed approach demonstrates promising results even in the presence of label noise.