计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (13): 266-275.DOI: 10.3778/j.issn.1002-8331.2309-0228

• 图形图像处理 • 上一篇    下一篇

改进的DeeplabV3Plus高分辨率遥感影像土地覆盖分类

朱凡,罗小波   

  1. 1.重庆邮电大学 计算机科学与技术学院,重庆 400065
    2.重庆邮电大学 空间大数据智能技术重庆市工程研究中心,重庆 400065
  • 出版日期:2024-07-01 发布日期:2024-07-01

Improved Land Cover Classification of DeeplabV3Plus High-Resolution Remote Sensing Imagery

ZHU Fan, LUO Xiaobo   

  1. 1.School of Computer Sciences and Technology, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
    2.Chongqing Engineering Research Centre for Spatial Big Data Intelligence Technology, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
  • Online:2024-07-01 Published:2024-07-01

摘要: 高分辨率遥感图像中提取的土地覆盖信息在城市规划建设和土地利用等领域具有巨大的价值,但由于土地覆盖类型复杂、光谱差异性较小等因素,目前对土地覆盖类别进行高质量的语义分割仍然受到一定限制。因此,针对该问题提出了一种新颖的全连接网络MFC-Net,该模型采用全新的基于点积注意力的空洞空间金字塔池化模块(DPA-ASPP),提高了聚合上下文信息方面的能力及效率。更进一步的,针对不同尺度的特征提出了注意力增强融合模块(AEFM)来增强特征表示,改善不同形状和大小地物的分割效果。该模型充分利用了高分辨率遥感影像中丰富的上下文信息及多尺度信息,在LoveDA大型遥感图像数据集上取得了优于当前主流模型的分割结果(84.26%MPA和69.67%MIOU)。

关键词: 遥感图像, 深度学习, 语义分割, 土地覆盖分类, LoveDA数据集

Abstract: Land cover information extracted from high-resolution remote sensing images is of great value in the fields of urban planning and construction and land use, etc. However, high-quality semantic segmentation of land cover categories is still restricted at present due to the complicated land cover types and small spectral variability. Therefore, a novel fully-connected network MFC-Net is proposed in this paper to address this challenge, and the model adopts a novel dot-product attention-atrous spatial pyramid pooling (DPA-ASPP), which improves the capability and efficiency in terms of aggregating contextual information. Furthermore, the attention enhanced fusion module (AEFM) is proposed for different scales of features to enhance the feature representation and improve the segmentation of features of different shapes and sizes. The model makes full use of the rich contextual and multi-scale information in high-resolution remote sensing images, and achieves segmentation results (84.26% MPA and 69.67% MIOU) better than the current mainstream models on LoveDA large remote sensing image dataset.

Key words: remote sensing imagery, deep learning, semantic segmentation, land cover classification, LoveDA dataset