A New Max-Min Convolutional Network for Hyperspectral Image Classification

Convolutional neural networks (CNNs) are a noteworthy tool for the classification of hyperspectral images (HSIs). CNNs apply non-linear activation functions to learn data patterns. One of them is the rectified linear unit (ReLU), which is a piece-wise linear function with a value which is the input if positive and zero otherwise. As a result, it is computationally efficient and tends to show good convergence behaviour. Nonetheless, its performance suffers from the so-called dying ReLU effect. This is usually managed by introducing more convolution layers increasing the depth of the model followed by a ReLU non-linearity layer that may hamper the convergence of network and produce a low classification accuracy due to data degradation. In order to alleviate these issues and transmit more information after the activation layer from the convolutional block, this paper develops a new end-to-end supervised feature learning framework called MaxMin-CNN, which works with sub-cubes of the original HSI data and successively applies 3-D MaxMin convolutional filters to improve the discrimination ability of the obtained spectral-spatial features by doubling the feature maps over all the convolutional layers. The new model gradually increases the heterogeneity of high-level spectral-spatial features across the MaxMin convolutional layers, enhancing the performance of HSI classification and reducing the model depth while preserving the classification performance. In order to validate the model, we report experiments over three widely used HSI datasets: Indian Pines, University of Pavia and University of Houston. The results reveal that the proposed MaxMin-CNN achieves a classification comparable to state-of-the-art classification models.