【作者】 李茂；

【导师】 周德义； 齐江涛；

【作者基本信息】 吉林大学 ， 农业电气化与自动化， 2022， 硕士

【摘要】 作物的生长环境信息和生长发育情况快速感知是精准农业实施中最基本、关键的问题。玉米田的秸秆管理策略深刻影响着玉米的长势,有必要进行玉米秸秆还田情况和苗期玉米长势信息获取技术的研究。本文针对黑土区保护性耕作玉米田典型信息获取技术的需求,基于玉米田图像和光谱信息,应用机器学习模型和图像处理方法对采集到的信息进行分析,研究获取玉米田秸秆覆盖率、秸秆混埋效果的技术,分析不同田间环境对玉米长势的影响,构建基于多源时空数据的玉米长势预测模型,为玉米田管理和决策提供技术支持和参考。本文主要研究内容如下:(1)针对玉米田地表秸秆覆盖率快速准确获取的需求,研究了利用无人机低空遥感影像和机器学习算法检测秸秆覆盖率的方法。利用改进的U-Net神经网络建立了秸秆覆盖图像语义分割模型。将该模型与VGG11 U-Net(以VGG11主干作为U-Net提取模块的算法)、三角形算法(TRIANGLER)以及基于多色彩空间信息与支持向量机结合的算法进行了评价和比较。(2)针对秸秆混埋效果难以量化的问题,研究利用机器视觉技术实现对秸秆混埋效果自动、准确、快速的测量。首先,本文建立了玉米田地表秸秆混埋可见光图像数据集。然后,基于图像处理技术和网格计数法,提出了一种对秸秆混埋效果进行定量描述和表征的方法。最后,基于该方法,本文开发了Web应用程序和图形界面系统。(3)为探究不同环境下玉米的生长情况以及构建玉米苗情预测模型,本文设置了田间对比试验,绘制玉米田归一化植被指数时序曲线,研究不同的播种深度、土壤坚实度和秸秆覆盖率分别对玉米长势的影响。将3种机器学习算法(随机森林、岭回归和线性拟合)应用于三类输入数据,建立了9种长势预测模型。通过上述研究表明:基于改进的U-Net神经网络建立的秸秆覆盖图像语义分割模型获取玉米田秸秆覆盖率的平均绝对偏差为3.56%,相比其他算法效果最优,该模型可为玉米田地表秸秆覆盖率的快速获取提供技术支持。秸秆混埋效果获取技术的研究使得研究人员和从业人员能以电子化方式采集秸秆混埋效果信息,更科学、高效地处理、分析和存储实验数据,满足对玉米田信息统一采集和管理的需求。在苗情信息获取技术的研究中,本文得出播深是对苗期玉米生长影响最大的特征变量,其次分别是土壤坚实度和秸秆覆盖率。试验研究发现3 cm有利于玉米苗期前期发育,5 cm可能更有利于玉米中后苗期的生长;对比试验中,200 k Pa的低土壤坚实度和40%的秸秆覆盖率较为适宜。在构建的9种长势预测模型中,基于多源时空数据的随机森林模型的拟合效果最好,训练集和测试集拟合优度分别为0.9840、0.9854,均方根误差为2.7122。综上所述,本文对玉米田秸秆管理信息和玉米长势信息获取技术的研究,为优化玉米田秸秆管理策略、提高播种质量进而增加玉米产量提供了参考和技术支持。更多还原

【Abstract】 Crop growth environment information and rapid perception of growth are the most basic issues of precision agriculture.The straw management strategy of corn field has a profound impact on the growth of maize,so it is necessary to study the situation of returning corn straw returning and the information acquisition technology of maize seedling growth.In this paper,according to the demand of typical information acquisition technology of conservation tillage maize field in black soil area,the collected information was analyzed by using machine learning model and image processing method based on the image and spectral information of maize field.In this paper,the technology of obtaining straw coverage rate and straw incorporation effect in maize field was studied.The effects of different field environments on maize growth were analyzed and a maize growth prediction model based on multi-source spatiotemporal data was established.This paper provides technical support and reference for maize field management and decision making.The main contents of this paper are as follows:(1)In response to the demand for rapid acquisition technology of surface straw coverage in corn fields,a method for detecting straw coverage using UAV low-altitude remote sensing images and machine learning algorithms was studied.A semantic segmentation model of strawcovered images is established by using the improved U-Net neural network.The model is evaluated and compared with VGG11 U-Net(algorithm with VGG11 backbone as U-Net extraction module),triangle algorithm(TRIANGLER),and algorithm based on multi-color space information combined with support vector machines.(2)It is difficult to quantify the effect of straw incorporation.In this paper,machine vision technology is used to measure the effect of straw incorporation automatically,accurately and quickly.First,the visible light image dataset of corn straw incorporation was established in this study.Then,based on image processing technology and grid counting method,a quantitative description and characterization method of straw incorporation effect is proposed.Finally,based on this method,a web application and a graphical interface system are developed in this study.(3)In order to explore the growth of maize in different environments and build a prediction model for maize seedlings,this paper set up a field comparison experiment.The normalized vegetation index time series curve of the corn field was drawn to study the effects of different sowing depths,soil firmness and straw coverage on the growth of corn.Five machine learning algorithms(random forest,ridge regression and linear fitting)were applied to three types of input data,and 9 growth prediction models were established.The above research shows that the average absolute deviation of the straw coverage rate obtained by the semantic segmentation model based on the improved U-NET neural network3.56%.This model can provide technical support for the rapid acquisition of corn surface straw coverage rate.The research on the acquisition technology of straw incorporation effect enables researchers and practitioners to collect information on straw incorporation effect electronically,process,analyze and store experimental data more scientifically and efficiently,and meet the demand of unified collection and management of corn field information.In the research of seedling information acquisition technology,this study concluded that sowing depth was the characteristic variable that had the greatest impact on maize growth at seedling stage,followed by soil firmness and straw coverage.The results showed that 3 cm is conducive to the early development of maize seedlings,and 5 cm may be more conducive to the growth of maize in the middle and late seedling stages;in the comparative test,low soil firmness(200 k Pa)and straw coverage rate of 40% are more suitable.Among the 9 growth prediction models,the random forest model based on multi-source spatiotemporal data has the best fitting effect.To sum up,this paper studies corn straw management information and corn growth information acquisition technology,which provides technical support for optimizing corn straw management strategies and improving corn yield.更多还原