LH-UNet: A Lighted Histoformer-Encoded U-Net for Sea Ice Recognition With High-Resolution Remote Sensing Images

Abstract

The background information is complex for sea ice monitoring when using high-resolution remote sensing images, which may lead to a certain degree of difficulty in extracting sea ice information. Lighted histoformer-encoded u-shaped convolutional network (LH-Unet), a semantic segmentation neural network for sea ice fine recognition is proposed in this study. Initially, a histogram transformer block (HTB) in the histoformer model was integrated into the encoder to enhance the accuracy of sea ice recognition. Additionally, a ghost convolution enhanced by a triple attention block (GTB) was introduced, significantly reducing the number of parameters and computational load while also improving accuracy. Furthermore, the mean intersection over union (mIoU) of the LH-UNet network proposed in this study was 96.25%, and the number of parameters of the mentioned architecture is less than 1 M. Notably, LH-UNet surpasses the performance of prominent deep learning methods, including U-Net, PSPNet, HRNet, and SegFormer. The results suggest a reliable technical support for sea ice identification basing on high-resolution remote sensing. Moreover, this study provides a possibility for the monitoring and early warning of sea ice distribution.