Comparative Study on Different Dimension Reduction Methods in Remote Sensing Ground Object Recognition

Hyperspectral image (HSI) datasets have redundant information due to its high dimension and correlation among bands, which affects recognition accuracy. We use Gabor convolution combined with principal component analysis (PCA) to reduce dimension. Gabor convolution extracts spatial feature information and PCA reduces the dimension of the spectral domain. Thus, we provide two methods for dimension reduction. The first dimension reduction method is in the order of spectral domain dimension reduction, spatial feature extraction and channel dimension reduction (SGC). The other is in the order of spatial feature extraction, channel dimension reduction and spectral domain dimension reduction (GCS). Based on the three HSI datasets (Indian Pines, Tea Farm and Salinas), the above three methods are used for dimension reduction. And then we use machine learning models for recognition. The results show that compared with PCA, the recognition accuracy of SGC is improved by 10.03%, 3.98% and 6.78% on average and the recognition accuracy of GCS is improved by 19.66%, 5.94% and 8.53% on average. Meanwhile, the GCS method has stronger model generalization and robustness.