基于RSTCNN的小麦叶片病害严重度估计

以小麦叶片条锈病和白粉病为研究对象,针对同类型病害的不同严重度之间的图像颜色及纹理特征差异较小,传统方法病害严重度估计准确率不高的问题,提出一种基于循环空间变换的卷积神经网络(Recurrent spatial transformer convolutional neural network,RSTCNN)对小麦叶片病害进行严重度估计。RSTCNN包含3个尺度网络,并由区域检测子网络进行连接。每个尺度网络以VGG19作为基础网络以提取病害的特征,同时为了统一区域检测过程中前后特征图的维度,在全连接层前引入空间金字塔池化(Spatial pyramid pooling,SPP);区域检测子网络则采用空间变换(Spatial transformer,ST)有效提取尺度网络特征图中病害的注意力区域。小麦叶片病害图像通过每个尺度网络中卷积池化层得到的特征图,一方面可作为预测病害严重度类别概率的依据,另一方面通过ST进行注意力区域检测并将检测到的区域作为下一个尺度网络的输入,通过交替促进的方式对注意力区域检测和局部细粒度特征表达进行联合优化和递归学习,最后对不同尺度网络的输出特征进行融合再并入到全连接层和Softmax层进行分类,从而实现小麦叶片病害严重度的估计。本文对采集的患有条锈病和白粉病的小麦叶片图像结合数据增强方法构建病害数据集,实验验证了改进后的RSTCNN在3层尺度融合的网络对病害严重度估计准确率较佳,达到了95.8%。相较于基础分类网络模型,RSTCNN准确率提升了7~9个百分点,相较于传统的基于颜色和纹理特征的机器学习算法,RSTCNN准确率提升了9~20个百分点。结果表明,本文方法显著提高了小麦叶片病害严重度估计的准确率。

Estimation of Non-CO_2 GHGs Emissions by Analyzing Burn Severity in the Samcheok Fire,South Korea

This study was performed to estimate the emission of non-CO 2 greenhouse gases(GHGs) from biomass burning at a large fire area.The extended methodology adopted the IPCC Guidelines(2003) equation for use on data from the Samcheok forest fire gathered using 30 m resolution Landsat TM satellite imagery,digital forest type maps,and growing stock information per hectare by forest type in 1999.Normalized burn ratio(NBR) technique was employed to analyze the area and severity of the Samcheok forest fire that occurred in 2000.The differences between NBR from pre-and post-fire datasets are examined to determine the extent and degree of change detected from burning.The results of burn severity analysis by dNBR of the Samcheok forest fire area revealed that a total of 16,200 ha of forest were burned.The proportion of the area characterized by a 'Low' burn severity(dNBR below 152) was 35%,with 'Moderate'(dNBR 153-190) and 'High'(dNBR 191-255) areas were at 33% and 32%,respectively.The combustion efficiency for burn severity was calculated as 0.43 for crown fire where burn severity was 'High',as 0.40 for 'Moderate' severity,and 0.15 for 'Low' severity surface fire.The emission factors for estimating non-CO 2 GHGs were separately applied to CO 130,CH 4 9,NO x 0.7 and N 2 O 0.11.Non-CO 2 GHGs emissions from biomass burning in the Samcheok forest fire area were estimated to be CO 44.100,CH 4 3.053,NO x 0.238 and N 2 O 0.038 Gg.