基于MHSA+DeepLab v3+的无人机遥感影像小麦倒伏检测

倒伏是影响小麦产量和质量的重要因素之一，及时准确获取倒伏信息有利于小麦良种选育中的倒伏损失鉴定。本文以小麦灌浆期和成熟期两个生长阶段的可见光无人机遥感影像为依据，构建多生长阶段小麦倒伏数据集，通过在DeepLab v3+模型中添加不同的注意力模块进行比较分析，提出一种基于多头自注意力（MHSA）的DeepLab v3+小麦倒伏检测模型。试验结果表明，提出的MHSA+DeepLab v3+模型的平均像素精度(Mean pixel accuracy, mPA)和均交并比(Mean intersection over union, mIoU)，灌浆期分别为93.09%和87.54%，成熟期分别为93.36%和87.49%。与代表性的SegNet、PSPNet和DeepLab v3+模型相比，在灌浆期mPA提高了25.45、7.54、1.82个百分点和mIoU提高了36.15、11.37、2.49个百分点，在成熟期mPA提高了15.05、6.32、0.74个百分点，mIoU提高了23.36、9.82、0.95个百分点。其次，相比于CBAM和SimAM两种注意力模块，在灌浆期及成熟期基于多头自注意力的DeepLab v3+表现均为最优，在灌浆期其mPA和mIoU分别提高了1.6、2.07个百分点和1.7、2.45个百分点，成熟期提高了0.27、0.11个百分点和0.26、0.15个百分点。研究表明提出的改进的DeepLab v3+模型能够有效地捕获灌浆期和成熟期的无人机小麦遥感图像中的倒伏特征，准确识别不同生育期的倒伏区域，具有良好的适用性，为利用无人机遥感技术鉴定小麦倒伏灾害等级和良种选育等提供了参考。

Detection of Wheat Lodging in UAV Remote Sensing Images Based on Multi-head Self-attention DeepLab v3+

Lodging is one of the main factors, which affect the yield and quality of wheat. Timely and accurate acquisition of wheat lodging information is beneficial to cultivating fine varieties and identifying lodging losses in agricultural insurance. A multi-growth stage wheat lodging dataset was constructed based on the visible light UAV remote sensing images of the two growth stages of wheat: grain filling stage and mature stage. By adding different attention modules to the DeepLab v3+ model for comparative analysis, a DeepLab v3+ wheat lodging detection model based on multi-head self-attention was proposed to accurately detect the lodging areas during wheat growth. The experimental results showed that the mPA and mIoU of the proposed multi-head self-attention DeepLab v3+ model were 93.09%, 87.54% (the grain filling stage) and 93.36%, 87.49% (the mature stage), which were improved by 25.45, 7.54, 1.82 (mPA) and 36.15, 11.37, 2.49 (mIoU) percentage points at the grain-filling stage and outperformed by 15.05, 6.32, 0.74 (mPA) and 23.36, 9.82, 0.95 (mIoU) percentage points at the mature stage, compared with the representative SegNet, PSPNet and DeepLab v3+ models, respectively. Secondly, compared with the two attention modules of CBAM and SimAM, DeepLab v3+ based on multi-head self-attention performed the best in both the grain filling stage and the mature stage, and its mPA and mIoU were increased by 1.6, 2.07 and 1.7, 2.45 percentage points at the grain-filling stage and increased by 0.27, 0.11 and 0.26, 0.15 percentage points at the mature stage. The results showed that the improved DeepLab v3+ model captured the lodging features in the UAV remote sensing images of wheat at the grain filling and mature stages effectively and identified the lodging areas in different growth stages precisely, and it had good applicability. It provided a reference for the identification of wheat lodging disaster grades and breeding of improved varieties by using UAV remote sensing technology.