密植棉田籽棉低空成像分层特征与测产机理研究

In order to exactly and efficiently solve difficult problems for estimating yield of cotton field based on the existing technologies, we put forward a new estimation idea, which is based on layer discriminant of low altitude cotton images. Low altitude unmanned aerial vehicle (UAV) imaging and other modern test methods will be adopted to evaluate different cultivars and planting patterns of the aerial images and cotton boll material characteristics. The clustering analysis will be done according to the image features of different layer cotton bolls, such as color, brightness, opening, shape and connectivity. Then the key image features will be selected basing on the results of clustering analysis. After this, deep learning algorithms, such as Deep Belief Networks (DBN),Convolutional Neural Network(CNN) and Generative Adversarial Network(GAN) will be researched to accurately discriminate the layers for cotton bolls. On this foundation, to confirm the relationship between different layers cotton bolls material characteristics, their images features and cotton yield, so as to build the model for cotton yield estimation. To seek the best parameters combination, orthogonal test would be applied to research aerial height, aerial shot velocity, camera angle, exposure time, aperture parameters influences on the estimation. Based on this condition, wind speed, light, temperature and humidity and other environmental factors should be considered to revise and optimize the estimation model. Ultimately mechanism of cotton yield estimation based on low altitude imaging and hierarchical cotton boll images will be verified the discrimination methods, and the theoretical foundation for realization of low altitude UAV imaging cotton yield will be laid.

针对现有估产技术对大田密植模式下棉花产量无法高效精准估测的难题，提出一种基于籽棉低空成像分层判别的棉花产量估测思路。采用无人机低空成像等现代测试方法获取不同品种和不同密植模式下的航拍图像及棉朵物料特性，分析研究不同棉层棉朵的色泽、亮度、开合度、形状轮廓、连通和层叠等图像特征的聚类关系，筛选关键图像特征，并研究利用深度置信网络、卷积神经网络、生成式对抗神经网络等深度学习算法对棉朵进行准确分层判别。在此基础上，明确不同棉层棉朵的物料特性和图像特征与棉田产量之间的关系，建立估产模型。通过正交试验，研究航拍高度、航拍速度、拍摄角度、曝光时间、通光孔径等参数对估产效果的影响关系，寻求最佳工作参数组合，并分析研究风速、光照、温湿度等环境因子对模型估产结果的影响，进而对估产模型进行修正和优化，最终探明棉朵低空成像分层判别为基础的棉田估产机理，为实现基于无人机低空成像的棉田产量估测奠定理论基础。

针对现有估产技术对大田密植模式下棉花产量无法高效精准估测的难题，提出一种基于籽棉低空成像分层判别的棉花产量估测方法，主要成果如下：.（1）提出采用无人机成像的方式进行棉田种植面积估测，首先采用无人机搭载可见光相机，对棉田及周边地区进行成像，并计算图像植被指数和纹理等共36项特征；其次通过主成分分析算法综合分析各个类型特征的优势并筛选有效特征，再采用最大似然分类法识别棉田；最后计算棉田面积，棉田面积总误差为0.51%，其中脱叶棉田面积误差为0.72%，未脱叶棉田面积误差为0.32%；.（2）利用低空无人机(UAV)成像和深度卷积神经网络(DCNN)估算密植棉田产量。由一架无人机在5m高度拍摄棉田图像。在此基础上，提出了一种改进的DCNN模型，该模型将编码器-解码器重组和扩张性卷积应用于棉铃像元分割，称为CD-SegNet。通过线性回归分析，建立了棉铃像元率与棉花产量之间的关系。对4个棉田的估产进行了机收加权验证。结果表明，CD-SegNet优于其他测试模型，包括SegNet、支持向量机(SVM)和随机森林(RF)。棉田估产平均误差为6.2%。.（3）通过对不同棉层棉朵物料特性的分析及果节高度测量，确认分层点为25.99cm和43.17cm。通过迁移学习模式构建了棉朵图像分层判别模型，分层准确率达到81%。在棉朵分层判别的基础上通过支持向量机建立了产量估测模型，统计计算得到单行遮挡率为15%、双行遮挡率为20%。矫正前，单行估产相对误差平均值为21.07%，双行估产相对误差平均值为28.34%，矫正后，单行估产相对误差平均值为8.07%，双行估产相对误差平均值为9.34%。