基于深度学习的无人机遥感小麦倒伏面积提取方法

Extraction of Lodging Area of Wheat Varieties by Unmanned Aerial Vehicle Remote Sensing Based on Deep Learning

为及时准确地提取小麦倒伏面积,提出一种融合多尺度特征的倒伏面积分割模型Attention_U^(2)-Net。该模型以U^(2)-Net为架构,利用非局部注意力(Non-local attention)机制替换步长较大的空洞卷积,扩大高层网络感受野,提高不同尺寸地物识别准确率;使用通道注意力机制改进级联方式提升模型精度;构建多层级联合加权损失函数,用于解决均衡难易度和正负样本不均衡问题。Attention_U^(2)-Net在自建数据集上采用裁剪方式提取小麦倒伏面积,查准率为86.53%,召回率为89.42%,F1值为87.95%。与FastFCN、U-Net、U^(2)-Net、FCN、SegNet、DeepLabv3等模型相比,Attention_U^(2)-Net具有最高的F1值。通过与标注面积对比,Attention_U^(2)-Net使用裁剪方式提取面积与标注面积最为接近,倒伏面积准确率可达97.25%,且误检面积最小。实验结果表明,Attention_U^(2)-Net对小麦倒伏面积提取具有较强的鲁棒性和准确率,可为无人机遥感小麦受灾面积及评估损失提供参考。

In order to extract the lodging area timely and accurately,a lodging area extraction model,namely Attention_U^(2)-Net,was proposed.By integrating multi-scale features and based on U^(2)-Net,Attention_U^(2)-Net employed non-local attention mechanism to replace the hole convolution with large step size,expanded the receptive field of high-level network and improved the recognition accuracy of ground objects with different sizes,and utilized channel attention mechanism to improve the cascade mode and enhanced the accuracy.A multi-level joint weighted loss function was designed to balance the difficult and easy samples,and solve the challenge of imbalance between positive and negative samples.Patch-based pipelines were utilized to extract the lodging area.Experimental results on the self-built dataset showed effectiveness of Attention_U^(2)-Net.The precision rate was 86.53%,the recall rate was 89.42%,and the F1 value was 87.95%,respectively.Compared with FastFCN,U-Net,U^(2)-Net,FCN,SegNet and DeepLabv3,Attention_U^(2)-Net achieved the highest F1 value and showed strong robustness and extraction accuracy.Compared with the labeled area,the extracted area obtained by Attention_U^(2)-Net via cropping method was the closest one,and the accuracy rate of lodging area can reach 97.25%.Meanwhile,the false detection area of Attention_U^(2)-Net was the smallest among all models.Experimental results showed that Attention_U^(2)-Net had strong robustness and high segmentation accuracy,which can be utilized as a valuable reference for UAV remote sensing of wheat affected area and loss assessment.