Channel Attention Module for Segmentation of 3D Hyperspectral Point Clouds in Geological Applications

Abstract

We develop a Transformer-based model enhanced with a Channel Attention Module (CAM) to capture the inter-channel dependencies in 3D hyperspectral point cloud data for geological applications. We hypothesize that specific channels of hyperspectral data correspond to distinct mineral types, and therefore, exploiting the relationships among these channels is beneficial for our analysis. We evaluate our method using the newly released Tinto dataset, which consists of 3D hyperspectral point clouds featuring three different spectral ranges: LongWave Infrared (LWIR), ShortWave Infrared (SWIR), and Visible-Near Infrared (VNIR).We explore four different CAMs from various networks—SENet, ECANet, CBAM, and DANet—and successfully integrate them into a CNN-based model to enhance feature representation. We specifically tailor the channel attention to our use of 3D hyperspectral point cloud data. Our experiments demonstrate significant improvements in performance after incorporating the CAM into our backbone model, which draws inspiration from the Point Cloud Transformer architecture and Vector Self-Attention mechanism. These results highlight the potential for further research into enhancing classification accuracy using hyperspectral data in geological applications. The code will be released on https://github.com/aldinorizaldy/CAM-Transformer.