DCA-Unet: Enhancing small object segmentation in hyperspectral images with Dual Channel Attention Unet

Abstract

Hyperspectral image (HSI), with its high spectral resolution, captures extensive information across multiple wavelengths beyond the visible spectrum, enabling the recognition of intricate object details and features. This capability renders HSI indispensable in scientific research and engineering applications. Despite the efficacy of fully convolutional networks in processing remote sensing data, current methods face challenges in accurately segmenting small objects in HSI and delineating the boundaries of similar or adjacent objects. To address these limitations, we propose a novel DCA-Unet framework for HSI semantic segmentation. This framework leverages a dual-channel attention module to capture feature dependencies across both spatial and spectral channel dimensions, thereby enriching contextual information. Specifically, positional and channel attention modules are incorporated after each layer of the Unet encoder to enhance pixel-level representation and spectral inter-channel dependencies, respectively. The fused output of these attention modules further strengthens the feature representation of the Unet encoder. In the final output, Dice loss is employed to quantify the overlap between predicted and actual segmentations, while Focal loss is utilized to balance background samples, thus improving segmentation performance for small objects. Experimental results demonstrate that the proposed DCA-Unet framework excels in HSI semantic segmentation tasks, particularly in the accurate segmentation of small objects.