首页|基于改进HRNet的遥感影像冬小麦语义分割方法

基于改进HRNet的遥感影像冬小麦语义分割方法

扫码查看
原文链接
  • 万方数据
  • 维普
冬小麦在影像中呈现田块碎小且分布零散等空间特征,同时影像包含的复杂地物对冬小麦识别造成干扰,易出现识别精度低且边界分割模糊等问题.为及时准确获取大范围冬小麦空间分布信息,该研究以高分二号卫星影像作为数据源,提出一种CAHRNet(change attention high-resolution Net)语义分割模型.采用HRNet(high-resolution Net)替换ResNet作为模型的主干网络,网络的并行交互方式易获取高分辨率的特征信息;联合 OCR(object-contextual representations)模块聚合上下文信息,以增强像素点与目标对象区域的关联性;3)引入坐标注意力(coordinate attention)机制,使网络模型充分利用有效的空间位置信息,以保留分割区域的边缘细节,提高对分布零散、形状多变的冬小麦田块的特征提取能力.试验结果表明,在自制的高分辨率遥感数据集上,CAHRNet模型的平均交并比(mean intersection over union,MIoU)和像素准确率(pixel accuracy,PA)分别达到 81.72%和 97.08%,MIoU相较U-Net、DeepLabv3+分别提高了 9.09、2.44个百分点;PA相较 U-Net、DeepLabv3+分别提高 6.80、1.59个百分点,说明CAHRNet模型具有较高的分割识别精度,可为进一步准确获取冬小麦作物分布信息提供技术支撑.
Improved HRNet semantic segmentation model for remote sensing images of winter wheat
Timely and accurate acquisition of winter wheat distribution is beneficial to the estimation of grain yield and optimization of planting structure.However,the previous means of acquisition cannot fully meet the needs of large-scale applications in recent years.Deep learning can be expected to introduce into agricultural remote sensing,in order to obtain the spatial distribution of agricultural resources,and further realize the development of agricultural modernization.Taking the winter wheat as the target object,the high-resolution Gaofen-2 remote sensing images were utilized to extract the winter wheat over a large area.Winter wheat was scattered distribution with the different field shapes in the remote sensing images.At the same time,the complex features were contained in the remote sensing images,leading to the interference in the recognition of winter wheat.In this study,a CAHRNet semantic segmentation network model was proposed to enhance the recognition accuracy and fuzzy details of field edges in the spatial distribution of winter wheat.The improved model first replaced the backbone feature network from ResNet with a high-resolution network HRNet.A unique parallel structure was connected to the multiple layers of sub-networks.The multiple sub-networks were continuously interacted with each other to repeatedly fuse the feature generated by multi-scale sub-networks.The spatial feature was enriched to rapidly locate the target object.The network was then maintained to extract the high-resolution feature.Secondly,the CAHRNet model was used to extract the spatial distribution of the winter wheat.The OCR module was introduced to aggregate the contextual feature of the target object.The extracted feature was classified by the backbone network.The aggregation of all the object regions was then evaluated to divide the regions into different categories.The correlation between the regions and the pixel points was enhanced to clarify the approximate location of the target object.Lastly,coordinate attention was introduced to capture the remote dependencies and retain the precise location of the target object,according to the dual spatial orientations.The correlation between multiple pixel points was enhanced to more accurately locate the exact position of the target object.The coordinate attention enhanced the spatial features of the target object and then reduced the detail loss caused by the model during feature extraction.Much more details were obtained than before in the prediction.The CAHRNet network model enhanced the supervision of the spatial features of the target object.The spatial relationship between the winter wheat and background pixels was established to enhance the region-pixel correlation.At the same time,the contextual feature of individual pixel points was utilized to enhance the pixel-point to pixel-point attention.The edge details of the target object region were preserved to highlight the overall features of winter wheat.The segmentation accuracy was improved to realize the fine extraction of winter wheat spatial distribution.The experimental results show that the optimal evaluation was achieved in each index on the high-resolution remote sensing dataset,with the MIoU and PA reaching 81.72%and 97.08%,respectively.The MIoU performance was improved by 9.90 and 2.44 percentage points,respectively,compared with the typical U-Net and DeepLabv3+,while the PA model was improved by 6.80 and 1.59 percentage points,respectively.The finding can provide technical support for the further acquisition of winter wheat crop distribution.

deep learningsemantic segmentationremote sensing imagewinter wheatintelligent interpretation

李旭青、吴冬雪、王玉博、陈文博、顾会涛

展开 >

北华航天工业学院遥感信息工程学院,廊坊 065000

北京师范大学遥感科学国家重点实验室,北京 100875

中国科学院空天信息创新研究院遥感科学国家重点实验室,北京 100101

深度学习 语义分割 遥感影像 冬小麦 智能解译

河北省"三三三人才工程"项目河北省高等学校科学技术研究青年拔尖人才项目遥感科学国家重点实验室开放基金&&北华航天工业学院研究生创新资助项目北华航天工业学院研究生创新资助项目

C20221032BJ2020056OFSLRSS20230367-Y50G04-9001-22/23YKY-2022-56YKY-2023-66

2024

10.11975/j.issn.1002-6819.202308134

农业工程学报
中国农业工程学会

农业工程学报

CSTPCD北大核心
影响因子：2.529
ISSN：1002-6819
年,卷(期)：2024.40(3)
  • 25