Alternative Weed Control Methods during Grape Establishment in the United States Upper Midwest

A field experiment was conducted to determine the effects of three mulches (straw, landscape fabric, and woodchip) and a single spring herbicide application (combination of oryzalin, flumioxazin, and glyphosate) on annual weed control and grapevine establishment in North Dakota. Vine growth, bud-break timing, bud hardiness and soil conditions were monitored to determine weed control method effects on vine progress. More consistent yellow foxtail (Setaria glauca L.) control occurred in mulched plots compared to plots treated with herbicide, particularly late in the season. In 2009, vines in mulched plots’ bud-break date were up to five days later when compared to vines grown in herbicide treated plots. However, no differences were observed in the spring of 2010. Overall, differences in growth rate were due to cultivar differences and not weed control methods. Results suggest that any of the three mulches could be used for annual weed control in northern vineyards during establishment as they offered at least as much weed control as the herbicide control and did not adversely affect vine establishment. However, continued research is needed to determine if mulches will alter fruit yield and quality upon vine maturation and together influence winter dieback of vines.

利用图像和机器学习检测大豆作物幼苗期玉米杂苗

在大豆-玉米轮作生产过程中,玉米杂苗会与大豆苗竞争水和肥料,而且很容易遮住大豆苗,影响害虫(如玉米根虫)的防控,降低大豆品质。因此,在大豆幼苗期及时检测出玉米杂苗并对其进行处理非常重要。传统的人工检测方法主观性强、效率低,传感器和算法的发展为自动检测玉米杂苗提供了更好的解决方案。本研究在温室环境下模仿田间条件,待玉米和大豆发芽后,连续5天用因特尔RealSense D435相机采集彩色图像,并人工裁剪幼苗图像区域,在此基础上对图像进行分割和去噪。在采集图像形状、色彩和纹理特征值后,对所采集的特征值进行权重分析,保留前10种重要的特征值导入基于特征的机器学习算法中进行模型训练和预测。预测结果表明,支持向量机模型(SVM)、神经网络(NN)和随机森林(RF)的预测精度分别为85.3%,81.5%和82.6%。将数据集导入GoogLeNet和VGG-16两种深度学习模型进行训练,预测精度分别为96.0%和96.2%。VGG-16模型在区分大豆幼苗和玉米杂苗中有较好的表现,彩色图像和VGG-16模型组成的系统可以自动检测大豆生长过程中玉米杂苗的情况,为农民提供准确的信息,帮助其进行生产决策和田间管理。