改进U-Net模型的保护性耕作田间秸秆覆盖检测

Straw coverage detection of conservation tillage farmland based on improved U-Net model

为了适应保护性耕作秸秆还田监测的技术需求,提出了一种改进的U-Net语义分割算法对秸秆覆盖率进行检测。首先,提出一种新的卷积模块代替原始U-Net框架中的卷积模块;其次,改进Inception结构,引入条纹池化和高效空间金字塔空洞卷积模块,形成新的Gception结构;最后,在模块中引入注意力机制。利用无人机采集田间地表图像,将改进的U-Net模型应用于自标注田间秸秆图像分割,与U-Net,PSP-Net,Link-Net,Res-Net,DSRA-Unet和DE-GWO算法进行对比实验,得到的平均交并比为80.05%,平均像素精确度为91.20%,覆盖率平均误差为0.80%。实验结果表明,改进U-Net模型的分割结果优于对比算法,能够保证特征提取的有效性和全局特征的完备性,有效剔除树影以及田内其他干扰因素。该模型适用于含有农机和树影等干扰的田间复杂场景,在大尺度图像中亦可获得较好的分割效果,可为大面积秸秆覆盖率检测提供技术支持。

To meet the technical requirements of straw return monitoring in conservation tillage,an improved U-Net semantic segmentation algorithm was proposed to detect straw coverage rate.First,a novel convolution module was developed to replace the convolution module in the original U-Net framework;Second,the inception structure was improved,and stripe pooling and efficient spatial pyramid hollow convolution modules were introduced for the design of a new Gception structure.Finally,an attention mechanism was introduced into this module.The effectiveness and progressiveness of the optimized U-Net model were evaluated on farmland surface images collected by a drone,and comparative experiments were conducted with the U-Net,PSP-Net,Link-Net,Res-Net,DSRA-Unet,and DE-GWO algorithms,obtaining mean intersection over union of 80.05%,an average pixel accuracy of 91.20%,and an average coverage error of 0.80%.The experimental results demonstrate that the segmentation capability of the im‐proved U-Net model was superior to those of the other algorithms,indicating the effectiveness of feature extraction and the integrity of global features,by effectively eliminating tree shadows and other interference factors over the farmland.These results prove that the proposed model achieves better segmentation not only in complex farmland scenes with agricultural machinery and tree shadow interference but also in large-scale images.Moreover,the model provides efficient algorithms for large-area straw coverage detection and other related image detection methods.