Hybrid Loss and 3D Convolutional Multilayer Perceptron for Hyperspectral Images Classification with Noisy Labels

Deep learning has been successfully applied to hyperspectral image classification, where large-scale datasets play a crucial role. However, accurately and comprehensively annotating these datasets is challenging. Although crowdsourcing companies and online annotation can accelerate the labeling process, they often come with label noise, while manual labeling is costly. Therefore, this paper proposes a Hybrid Loss Function and 3D Convolutional Multilayer Perceptron Model (HDCMM). Firstly, spatial and spectral features are simultaneously extracted using a three-dimensional convolutional neural network (3DCNN), fully utilizing the 3D information of hyperspectral images. Then, a Multilayer Perceptron (MLP) is used for classification decisions to fully exploit its advantages in handling high-dimensional feature spaces. To further enhance the model's robustness to noisy labels, a hybrid loss function incorporating Reverse Cross-Entropy and Normalized Cross-Entropy is introduced. These two loss functions perform excellently in handling noisy labels, effectively reducing the interference of noise in model training. By combining the powerful feature extraction capabilities of the 3D CNN and the excellent classification performance of the MLP with the hybrid loss function, the proposed method demonstrates superior performance in hyperspectral image classification tasks with noisy labels. Experimental results on the Pavia Center (PC) dataset show that HDCMM outperforms other deep learning based models in classifying HSI with noisy labels.