Spatial-Spectral Fusion Network for Plastic Classification Using Hyperspectral Image

Plastic plays a critical role in modern life and industry due to its unique properties and wide range of applications. Accurate classification of plastics enhances resource utilization and reduces environmental impact. In this work, we propose a spatial-spectral fusion network (SSF-Net) to achieve plastic classification. First, we collect an HSI dataset with five classes of plastics. Second, we introduce a spatial-spectral dilated convolution (SSDConv), which integrates multi-scale features with different dilation rates, to capture the hierarchical details of HSIs. Third, we develop a spatial-spectral attention module (SSAM) to extract critical information from HSIs through both spatial and spectral perspectives. Finally, we construct SSF-Net by combining SSDConv and SSAM to facilitate plastic classification. Experimental results on our collected HSIs demonstrate that SSF-Net achieves superior classification performance compared to multiple hyperspectral classification networks and traditional classification methods.