Unsupervised clustering for intrinsic mode functions selection in Hyperspectral image classification

In the realm of hyperspectral image classification, traditional methods typically eliminate spectrum noise to enhance spectral features, followed by the application of supervised techniques to improve classification efficiency. However, the use of ensemble empirical model decomposition (EEMD) has gained attention in recent years for its ability to select intrinsic mode functions (IMFs) and reconstruct a new spectrum. Nevertheless, concerns arise regarding the potential suboptimality of selected IMFs and their impact on classification accuracy. To address this issue, our study leverages EEMD to decompose each substance's spectrum into multiple IMFs, which are then clustered using K-means and hierarchical clustering. The proposed unsupervised clustering approach combines IMFs with similar features to create a new spectrum. Notably, our model surpasses the limitations of suboptimal IMF selection, leading to enhanced classification accuracy. Extensive experiments were conducted on hyperspectral data contaminated with high noise signals. The evaluation metrics employed encompassed accuracy as the primary measure. Our model demonstrated superior performance, achieving a significant improvement in accuracy from 0.6640 to 0.9177 compared to previous approaches. In conclusion, our proposed model introduces advancements by incorporating EEMD and unsupervised clustering techniques. The experimental results substantiate its superiority in achieving higher classification accuracy, overcoming the limitations of traditional methods. This study contributes to the field of hyperspectral image classification by offering an effective solution that addresses the challenges posed by suboptimal IMF selection.