基于深度学习的多光谱卫星遥感图像地物分类算法通常选用RGB波段而忽略NIR等波段数据，其网络的特征提取与应用扩展能力有待提升。针对这一问题，文章提出一种基于改进YOLOv5 的多光谱卫星遥感图像地物分类方法(即VN-YOLOv5-Seg网络)，该方法联合RGB与NIR波段数据作为输入，以YOLOv5 目标检测网络作为骨干网络，使用ProtoNet网络作为分割头将目标检测转换为像素级的地物分类任务。为了验证VN-YOLOv5-Seg网络的有效性，文章选用GID-15 数据集，分别使用RGB波段、RGB+NIR波段作为网络输入进行试验，并将VN-YOLOv5-Seg与其他地物分类网络的分类结果进行对比分析。试验结果表明，在RGB波段基础上引入NIR波段，平均交并比(Mean Intersection over Union，mIoU)提高了2。5%;相较于FCN分割头，mIoU提升了8。1%;相较于PSPNet、DeepLabV3 和U-Net方法，mIoU分别提高了 2。6%、1。2%和 1。4%。试验结果充分验证了方法的有效性，以及引入更多波段信息用于地物分类的必要性。
Landcover Classification Method for Multispectral Satellite Remote Sensing Imagery Based on Improved YOLOv5
Deep learning-based algorithms of land cover classification for multispectral satellite remote sensing imagery classification typically utilize RGB band data while overlooking other bands such as NIR,and the networks'feature extraction and application expansion capabilities need improvement.Regarding this issue,this paper proposes a land cover classification method multispectral satellite remote sensing imagery classification method based on the improved YOLOv5,called VN-YOLOv5-Seg.This method jointly utilizes RGB and NIR band data as inputs,adopts YOLOv5 object detection network as the backbone network,and employs the ProtoNet network as the segmentation head to convert object detection into pixel-level land cover classification tasks.The GID-15 dataset is used for experiments to verify the effectiveness of this method,with RGB band and RGB+NIR band as network inputs.Comparative analyses are conducted between VN-YOLOv5-Seg and other land cover classification networks.Experimental results demonstrate that by adding the NIR band to the RGB band,the mean Intersection over Union(mIoU)is improved by 2.5%.Compared to the FCN segmentation head,the mIoU is improved by 8.1%.Compared to PSPNet,DeepLabV3,and U-Net methods,the mIoU is increased by 2.6%,1.2%,and 1.4%respectively.These results fully validate the effectiveness of the method and the necessity of introducing more band information for land cover classification.
multispectral imagerylandcover classificationobject detection networkNIR band
NETL
NSTL
万方数据
