Hyperspectral Sea Ice Image Classification Based on the Spectral-Spatial-Joint Feature with the PCA Network

Abstract

Sea ice is one of the most prominent causes of marine disasters occurring at high latitudes. The detection of sea ice is particularly important, and the classification of sea ice images is an important part of sea ice detection. Traditional sea ice classification based on optical remote sensing mostly uses spectral information only and does not fully extract rich spectral and spatial information from sea ice images. At the same time, it is difficult to obtain samples and the resulting small sample sizes used in sea ice classification has limited the improvement of classification accuracy to a certain extent. In response to the above problems, this paper proposes a hyperspectral sea ice image classification method involving spectral-spatial-joint features based on the principal component analysis (PCA) network. First, the method uses the gray-level co-occurrence matrix (GLCM) and Gabor filter to extract textural and spatial information about sea ice. Then, the optimal band combination is extracted with a band selection algorithm based on a hybrid strategy, and the information hidden in the sea ice image is deeply extracted through a fusion of spectral and spatial features. Then, the PCA network is designed based on principal component analysis filters in order to extract the depth features of sea ice more effectively, and hash binarization maps and block histograms are used to enhance the separation and reduce the dimensions of features. Finally, the low-level features in the data form more abstract and invariant high-level features for sea ice classification. In order to verify the effectiveness of the proposed method, we conducted experiments on two different data collection points in Bohai Bay and Baffin Bay. The experimental results show that, compared with other single feature and spectral-spatial-joint feature algorithms, the proposed method achieves better sea ice classification results (94.15% and 96.86%) by using fewer training samples and a shorter training time.