The deployment of convolutional neural networks (CNNs) to classify hyperspectral images is extensively discussed in the research study. A number of different algorithms and approaches are applied, including 2-D CNN, 3-D CNN, support vector machine (SVM), regression models, and other state-of-The-Art deep learning models, although these methods do not show good performance for hyperspectral image classification. Furthermore, 3-D CNNs require a lot of computational power and are not mainly employed, whereas 2-D CNNs do not constitute multiresolution image processing and exclusively focus on spatial features. However, 3D-2D CNNs aim to incorporate spectral and spatial features, and their efficiency while being evaluated on various datasets tends to be limited. Moreover, a number of deep learning models have been proposed recently, but their performance is still limited. In order to solve these problems, in this article, we propose a novel deep hyperspectral shot, a deep smooth wavelet CNN shots ensemble for hyperspectral image classification. A deep smooth wavelet CNN utilizes layers of wavelet transform to extract spectral features. The computation of a wavelet transform is less intensive as compared to the computation of a 3-D CNN. After that, the extracted spectral features are integrated into 2-D CNN, which generates spatial features, as a result, generates a spatial-spectral feature vector for classification. Furthermore, we introduce the snapshots generation method and employ a cyclic annealing schedule to converge to several local minima along its optimization path and save the models. We build several snapshots of the deep hyperspectral shots model to enhance the performance of our proposed method. We propose the snapshots optimization and ensemble selection approach in order to solve the optimization problem within ensemble creation and further enhance the performance. In addition, we also introduce a novel activation function called Relish to increase spatial-spectral feature propagation and advance for smoother gradients. Overall, we ensemble the snapshots of our proposed method and achieved that can classify multiresolution HSI data with high accuracy. Experiments performed on benchmark datasets, our proposed method, deep hyperspectral shots, achieved overall accuracies of 99.96%, 97.91%, and 99.49% on the Salinas, Indian Pines, and Pavia University datasets against the state-of-The-Art methods.
Deep Hyperspectral Shots: Deep Snap Smooth Wavelet Convolutional Neural Network Shots Ensemble for Hyperspectral Image Classification
Pau, Giovanni
2024-01-01
Abstract
The deployment of convolutional neural networks (CNNs) to classify hyperspectral images is extensively discussed in the research study. A number of different algorithms and approaches are applied, including 2-D CNN, 3-D CNN, support vector machine (SVM), regression models, and other state-of-The-Art deep learning models, although these methods do not show good performance for hyperspectral image classification. Furthermore, 3-D CNNs require a lot of computational power and are not mainly employed, whereas 2-D CNNs do not constitute multiresolution image processing and exclusively focus on spatial features. However, 3D-2D CNNs aim to incorporate spectral and spatial features, and their efficiency while being evaluated on various datasets tends to be limited. Moreover, a number of deep learning models have been proposed recently, but their performance is still limited. In order to solve these problems, in this article, we propose a novel deep hyperspectral shot, a deep smooth wavelet CNN shots ensemble for hyperspectral image classification. A deep smooth wavelet CNN utilizes layers of wavelet transform to extract spectral features. The computation of a wavelet transform is less intensive as compared to the computation of a 3-D CNN. After that, the extracted spectral features are integrated into 2-D CNN, which generates spatial features, as a result, generates a spatial-spectral feature vector for classification. Furthermore, we introduce the snapshots generation method and employ a cyclic annealing schedule to converge to several local minima along its optimization path and save the models. We build several snapshots of the deep hyperspectral shots model to enhance the performance of our proposed method. We propose the snapshots optimization and ensemble selection approach in order to solve the optimization problem within ensemble creation and further enhance the performance. In addition, we also introduce a novel activation function called Relish to increase spatial-spectral feature propagation and advance for smoother gradients. Overall, we ensemble the snapshots of our proposed method and achieved that can classify multiresolution HSI data with high accuracy. Experiments performed on benchmark datasets, our proposed method, deep hyperspectral shots, achieved overall accuracies of 99.96%, 97.91%, and 99.49% on the Salinas, Indian Pines, and Pavia University datasets against the state-of-The-Art methods.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.