基于深度学习的5种树皮纹理图像识别研究

目的针对在树皮图像识别时,现有的算法和识别过程过于复杂的问题,提出了基于深度学习的方法来对不同树种的树皮图像进行识别。方法本文以5种常见树种的树皮纹理图像为例,采用基于卷积神经网络的深度学习方法,将原始图像直接作为输入,通过卷积和池化层对图像的低级、高级特征进行自动提取,解决了手动提取纹理特征的困难和问题;在此基础上,对CNN模型结构进行改进,采用带Maxout的ELU激励函数来代替ReLU函数,解决模型的偏移和零梯度问题;对损失函数进行改进,通过添加规范项来优化结构参数,并使用分段常数衰减法对学习率进行动态调控;最后采用softmax分类器对图像类别进行输出。结果对5个树种的树皮图像共计10 000张图像进行实验,其中每类选取200张图像作为测试集。最终训练准确率达到93.80%,测试集识别准确率为97.70%。另外,为验证本文方法的可行性,与传统人工特征提取法,提取HOG特征、Gabor特征和灰度共生矩阵统计法,训练SVM分类器。通过实验比较,本文方法识别准确率最高。结论本文提出的基于深度学习的树皮纹理图像识别方法是可行的,提高了识别效率和精度,为树种的智能化识别提供新的参考。      

基于局域最大值法单木位置探测的适宜模型研究

以凉水自然保护区为研究区域,基于机载激光雷达数据,采用动态窗口局部最大值法对郁闭度较高的针叶林进行单木位置自动提取。采用树冠高度模型(CHM)和树冠最大模型(CMM)配合两种动态窗口,即树高--树冠大小回归方程和该方程的95%预测下限来探测树冠顶点,用探测百分比、1∶1对应关系的单木个数、生产者精度和用户精度进行了精度评价。结果表明:利用CMM能够抑制树冠内部枝杈产生的错判现象;利用树高--树冠大小回归方程95%的预测下限做动态窗口,能够有效防止在局部最大值方法中产生的小树漏测现象。因此,利用CMM和95%的预测下限做动态窗口的局域最大值法有利于提高单木位置探测的精度,为密林中自动地探测单木位置提供依据。      

基于多特征融合的花卉种类识别研究

花卉种类识别作为植物自动分类识别的重要分支,有着很高的研究和应用价值。针对当前花卉特征描述存在的局限和花卉识别准确率较低的实际情况,以花卉图像为研究对象,首先对复杂背景图像采用基于显著性检测的Grab Cut分割算法进行预处理,得到单一背景图像;然后在提取花卉图像花冠(所有花瓣)颜色和形状特征的基础上,创新性地提取花蕊区域的颜色和形状所包含的特征信息,并将提取到的18个特征融合成单一特征向量。以支持向量机(SVM)算法为基础构建分类器,通过实验确定核函数与最佳参数;对360幅自建花卉样本库(24个种类,每个种类15幅)进行训练和测试,其中240幅作为训练样本,120幅作为测试样本,并与基于不同特征组合的识别方法进行比较。结果表明:本文提出的基于多特征融合的识别方法具有较高的识别准确率,识别率可以达到92.50%。对通用花卉样本库Oxford 17 flower进行训练与测试,选取其中340幅作为训练样本,170幅作为测试样本,取得了较好的识别效果,验证了本文方法的有效性。      

Evaluation of an algorithm for automatic detection of broad-leaved weeds in spring cereals

Lack of automatic weed detection tools has hampered the adoption of site-specific weed control in cereals. An initial object-oriented algorithm for the automatic detection of broad-leaved weeds in cereals developed by SINTEF ICT (Oslo, Norway) was evaluated. The algorithm (“WeedFinder”) estimates total density and cover of broad-leaved weed seedlings in cereal fields from near-ground red–green–blue images. The ability of “WeedFinder” to predict ‘spray’/‘no spray’ decisions according to a previously suggested spray decision model for spring cereals was tested with images from two wheat fields sown with the normal row spacing of the region, 0.125 m. Applying the decision model as a simple look-up table, “WeedFinder” gave correct spray decisions in 65–85% of the test images. With discriminant analysis, corresponding mean rates were 84–90%. Future versions of “WeedFinder” must be more accurate and accommodate weed species recognition.     

计算机辅助小麦图像识别应用中颜色特征基本参量的表达

在计算机辅助图像处理中，运用颜色特征进行图像的分类和识别是简便而有效的一种方法。然而，颜色特征的表达和提取是否准确、合理直接决定着分类和识别的可靠性。本文在重点分析RGB、HIS和L*a*b*三种常用颜色模式基本参量含义及相互间关系的基础上，结合小麦图像自身的特点，通过对30幅小麦图像在三种颜色模式下的9个基本参量进行主成分分析，建立了应用于小麦图像识别的颜色特征基本参量表达式，并对这三种颜色模式的9个基本参量进行了分类，提出了确定而有意义的表征小麦颜色特征的主成分指标。结果如下：基于第一主成分的分类指标综合表达出小麦冠层的亮绿色特点，分类结果具有较高的准确性和可靠性；第二主成分指标主要表达小麦冠层黄绿颜色变化的特点，能够形成连续的量化指标空间。第三主成分指标主要表达小麦正常绿色的情况，在图像获取亮度差异较小时可以进行小麦正常绿色值的评价。     

Investigation on pistachio distribution in the mountain regions of northeast Iran by ALOS

Iran supports five different vegetation zones. One of those is the Irano-Touranian zone that is located in the northeast of Iran. This vegetation zone includes arid and semi-arid lands, and its area is about 3.5 million hm². It supports growth of pistachio (Pistacia vera), a deciduous-broadleaved species, which is one of the ecologically and economically most important native species. In this study, we analyzed three images acquired by ALOS satellite, including 10m resolution multispectral band (AVNIR-2), 2.5 m resolution “Backward” PRISM image, and 2.5 m resolution “Nadir” PRISM image, based on a provided rational polynomial coefficient (RPC). Using the “Backward” and “Nadir” images, a 2.5 m resolution digital elevation model (DEM) was produced. Four methods with AVNIR-2 and PRISM data were used to produce pan-sharpening images and conduct an object-based feature extraction process. Normalized Difference Vegetation Index (NDVI) was used to determine the maximum distribution of pistachio in related elevation. The accuracy of the DEM was tested on 28 ground control points in the pair image as tie points, with the value of parallax error of 0.9027 m. The created elevation map indicated that pistachio trees grow up at 650m above sea level (a.s.l.). The result from NDVI in the related elevation showed the maximum density of pistachio at 800m a.s.l. In addition, the result of feature extraction in the forest showed the area of each target element calculated. The results of this research will improve decision-making and lead to sustainable management in general.     

基于无人机影像阴影去除的苹果树冠层氮素含量遥感反演

【目的】去除无人机多光谱遥感影像中的阴影,以提高苹果树冠层氮素含量反演模型精度。【方法】以山东省栖霞市苹果园为试验区,利用2019年6月采集的无人机多光谱影像,分别基于归一化阴影指数（normalized shaded vegetation index,NSVI）和归一化冠层阴影指数（normalized difference canopy shadow index,NDCSI）去除果树冠层多光谱影像中的阴影,提取非阴影区域果树冠层光谱信息;通过相关性分析方法,将基于原始光谱影像和基于NSVI、NDCSI去除阴影后提取的光谱数据与实测叶片氮素含量进行相关性分析,分别筛选氮素含量的敏感波段并构建光谱参量;采用偏最小二乘（partial least square,PLS）及支持向量机（support vector machine,SVM）方法构建果树冠层氮素含量反演模型并进行精度检验。【结果】绿光波段和红光波段为果树冠层氮素含量反演的敏感波段;阴影削弱了果树冠层的光谱信息,去除阴影前后,冠层多光谱各波段光谱差异较大,在红边波段及近红外波段尤为明显;基于2个阴影指数去除阴影后构建的氮素反演模型精度均有提升,最优模型为基于NDCSI去除阴影后构建的支持向量机氮素含量反演模型,该模型建模集R²和RPD分别为0.774、1.828;验证集R²和RPD分别为0.723、1.819。【结论】基于NDCSI可有效去除无人机多光谱果树冠层影像中的阴影,提高氮素含量反演精度,为果园氮素精准管理提供了有效参考。     

基于LiDAR数据的单木提取算法研究

有效了解森林生长变化信息对森林资源的保护以及生态环境的研究具有重要意义。近年来,激光雷达数据是森林清查中主要的LiDAR数据源。但是现有机载LiDAR数据单木分割算法在密集林区应用精度较低,尤其在中下层冠层单木提取精度不高,存在漏检的现象。针对以上问题,采用归一化割（normalized cut,Ncut）方法直接对激光点云数据进行初始分割,然后设定冠层的形状参数与点云数量阈值,利用全局最大值重复Ncut方法,对林区下层冠层进行探测,实现单木的精确提取。结果表明,与只利用归一化割方法提取单木结果相比,本研究方法使单木提取误判率由22.66%降至3.9%,识别率由原先的68.49%提升至86.63%,有效规避了上层冠木对下层遮盖导致的下层单木漏检情况,提高了在冠层中间层和下层树木的识别率,可为今后森林清查、森林资源管理提供分割方法的选择,也为森林分类、单木分割提供样例。     

光照条件和果台年龄对秦冠苹果商品品质的影响

通过光照条件和果台年龄对果实商品品质影响的调查表明，苹果商品品质主要取决于冠内光照条件，果台年龄的作用甚小，果实大小，干物质一，可溶性固形物含量及果面红色覆盖率与光照条件呈密切正相关，而果实含酸量和硬度则与光照水平呈负相关，从果实大小考虑，２０％￣３０％ＦＳ的光照环境可能为秦冠苹果生产的下限。     

A procedure for estimating the number of green mature apples in night-time orchard images using light distribution and its application to yield estimation

A procedure for estimating the number of mature apples in orchard images captured at night-time with artificial illumination was developed and its potential for estimating yield was investigated. The procedure was tested using four datasets totaling more than 800 images taken with cameras positioned at three heights. The procedure for detecting apples was based on the observation that the light distribution on apples follows a simple pattern in which the perceived light intensity decreases with the distance from a local maximum due to specular reflection. Accordingly, apple detection was achieved by detecting concentric circles (or parts of circles) in binary images obtained via threshold operations. For each dataset, after calibration of the procedure using 12 images, the estimates of the number of apples were within a few percent of the number of apples counted by visual inspection. Yield estimations were obtained via multi-linear models that used between two and six images per tree. The results obtained using all three cameras were only slightly better than those obtained using only two cameras. Using images from only one side of the tree did not worsen the results significantly. Overall, the yield estimated by the best models was within \(\pm\)10 % of the actual yield. However, the standard deviation of the yield estimation errors corresponded to ~26–37 % of the average tree yield, indicating that improvements are still needed in order to achieve accurate yield estimation at the single-tree level.     

基于透射光和反射光图像同位分割的小麦品种识别方法研究

为提高基于机器视觉的小麦品种识别准确性,本文通过透射光和反射光同位图像分割对种子颜色特征参数进行了优化提取.采用透射光图像辅助反射光图像分割的方式从种子图像中分割出胚部区域,并分别提取小麦整粒、种胚、胚乳区域的颜色特征参数.以济麦22、济麦44、京麦9、京麦11共4个品种种子作为研究对象,利用HALCON机器视觉软件获...     

Within-field nitrogen response in corn related to aerial photograph color

Precise management of nitrogen (N) using canopy color in aerial imagery of corn (Zea mays L.) has been proposed as a strategy on which to base the rate of N fertilizer. The objective of this study was to evaluate the relationship between canopy color and yield response to N at the field scale. Six N response trials were conducted in 2000 and 2001 in fields with alluvial, claypan and deep loess soil types. Aerial images were taken with a 35-mm slide film from ≥1100 m at the mid- and late-vegetative corn growth stages and processed to extract green and red digital values. Color values of the control N (0 kg N ha⁻¹) and sufficient N (280 kg N ha⁻¹ applied at planting) treatments were used to calculate the relative ratio of unfertilized to fertilized and relative difference color values. Other N fertilizer treatments included side-dressed applications in increments of 56 kg N ha⁻¹. The economic optimal N rate was weakly related (R ² ≤ 0.34) or not related to the color indices at both growth stages. For many sites, delta yield (the increase in yield between control N and sufficient N treatments) was related to the color indices (R ² ≤ 0.67) at the late vegetative growth stage; the best relationship was with green relative difference. The results indicate the potential for color indices from aerial photographs to be used for predicting delta yield from which a site-specific N rate could be determined.     

基于HSV空间的玉米果穗性状的检测

为高效检测玉米果穗性状,建立了基于HSV(色调、饱和度、明度值)空间的玉米果穗性状的检测方法:使用机器视觉技术采集绿色背景玉米果穗图像,用HSV直方图阈值算法去除绿色背景,用FFT滤波器去除尖锐边缘和噪声,运用粒子滤波分离单一图像中的多个玉米果穗图像,并采用形态学腐蚀方法,经过4次迭代腐蚀,得到玉米果穗中间3行;检测玉米果穗的大小、形状、纹理和颜色4个特征的性状。随机检测67张玉米果穗样本图像的结果表明,果穗大小和形状特征检测的准确率为100%,果穗颜色和纹理特征检测的准确率分别为98.55%和96.25%,平均每果穗检测时间为0.1 s。     

Lettuce calcium deficiency detection with machine vision computed plant features in controlled environments

Conventional greenhouse environmental conditions are determined by observation. However, destructive or invasive contact measurements are not practical for real-time monitoring and control applications. At the canopy scale, machine vision has the potential to identify emerging stresses and guide sampling for identification of the stressor. A machine vision-guided plant sensing and monitoring system was used to detect calcium deficiency in lettuce crops grown in greenhouse conditions using temporal, color and morphological changes of the plant. The machine vision system consisted of two main components: a robotic camera positioning system and an image processing module. The machine vision system extracted plant features to determine overall plant growth and health status, including top projected canopy area (TPCA) as a morphological feature; red-green-blue (RGB) and hue-saturation-luminance (HSL) values as color features; and entropy, energy, contrast, and homogeneity as textural features. The machine vision-guided system was capable of extracting plant morphological, textural and temporal features autonomously. The methodology developed was capable of identifying calcium-deficient lettuce plants 1 day prior to visual stress detection by human vision. Of the extracted plant features, TPCA, energy, entropy, and homogeneity were the most promising markers for timely detection of calcium deficiency in the lettuce crop studied.     

云南红豆杉单木生物量模型选择与树冠调控

拟合并选择云南红豆杉单木生物量模型,明确树冠结构调控技术。以3年生福建省明溪县云南红豆杉原料林为对象,采用4种自变量类型模型,并导入树冠结构特征变量拟合生物量模型;经配对t检验、偏差检验与估算预测精度,优选系列单木生物量模型。比较导入树冠结构特征前后模型的拟合与预测效果,明确并验证树冠结构调控技术。结果表明,采用D²H变量拟合的模型效果最优,其次为DH变量,然后是D²变量,最后是D、H变量。采用D²H变量较D、H变量明显提高模型的拟合与预测效果。导入树冠结构特征后明显提高模型拟合与预测效果。冠幅、冠形率显著影响红豆杉系列单株生物量,促进系列生物量的树冠结构调控重点与方向是促进冠幅宽大;经验证,树冠结构调控技术显著提高系列单木生物量是可行的。自变量类型显著影响模型的拟合与预测效果,选择自变量类型以优化红豆杉单木生物量模型是必要和可行的。导入树冠结构特征因子明显提高单木生物量模型的拟合效果与预测精度。选出4个最优红豆杉单木生物量模型,其决定系数不小于0.862,预测精度均不小于95.18%,M_APE均不大于16.95%,可以用于生产中的生物量预测。     

Seven samurai to protect “our” food: the reform of the food safety regulatory system in Japan after the BSE crisis of 2001

Using the case of food safety governance reform in Japan between 2001 and 2003, this paper examines the relationship between science and trust. The paper explains how the discovery of the first BSE positive cow and consequent food safety scandals in 2001 politicized the role of science in protecting the safety of the food supply. The analysis of the Parliamentary debate focuses on the contestation among legislators and other participants over three dimensions of risk science, including “knowledge,” “objects,” and “beneficiaries.” The metaphor of “seven samurai” and the relationally situated roles of “samurai,” “bandits,” and “beneficiaries” are used to show that in the process of policy making certain moral and ethical expectations on a new expert institution for food safety were contested and negotiated to frame responsibilities and commitments of social actors for creating the food system based on trust.

Delineating productivity zones in a citrus grove using citrus production,tree growth and temporally stable soil data

The productivity of a citrus grove with variation in tree growth was mapped to delineate zones of productivity based on several indicator properties. These properties were fruit yield, ultrasonically measured tree canopy volume, normalized difference vegetation index (NDVI), elevation and apparent electrical conductivity (EC_a). The spatial patterns of soil series, soil color and EC_a, and their correspondence with the variation in yield emphasized the importance of variation in the soil in differentiating the productivity of the grove. Citrus fruit yield was positively correlated with canopy volume, NDVI and EC_a, and yield was negatively correlated with elevation. Although all the properties were strongly correlated with yield and were able to explain the productivity of the grove, citrus tree canopy volume was most strongly correlated (r = 0.85) with yield, explaining 73% of its variation. Tree canopy volume was used to classify the citrus grove into five productivity zones termed as ‘very poor’, ‘poor’, ‘medium’, ‘good’ and ‘very good’ zones. The study showed that productivity of citrus groves can be mapped using various attributes that directly or indirectly affect citrus production. The productivity zones identified could be used successfully to plan soil sampling and characterize soil variation in new fields.     

基于支持向量机的路面图像分类方法

针对路面结构特征,提出一种颜色与纹理特征相融合并结合模糊支持向量机的路面分类识别方法。提取路面图像的HSV颜色空间的颜色矩作为颜色特征,采用灰度共生矩阵法提取纹理特征,融合路面图像的颜色特征与纹理特征,采用模糊支持向量机进行支持向量特征训练,通过训练得到能尽可能多的满足每一种图像的样本数据特征的特征向量。通过实验,对比了采用传统的支持向量机与模糊支持向量机对路面分类识别的正确率。实验表明本研究所提出方法的有效性。     

基于自动定向的苹果品质智能分级生产线设计

设计了一种基于自动定向原理的用于苹果品质动态、实时检测的智能化分级生产线,由苹果输送系统、自动定向小车、计算机视觉识别控制系统和分级执行装置组成。其中,苹果输送系统将苹果按分级节拍输送到自动定向小车上,由自动定向小车将苹果果梗花萼轴线定向到垂直于水平面的位置,位于圆周分布的3个摄像头一次性采集苹果表面信息,通过计算机识别控制系统进行智能识别,根据国家标准判断每个苹果的等级,并确定苹果的位置信息,通过计算机识别控制系统发出指令传输给分级控制装置,完成苹果的分级。     

The use and abuse of participatory rural appraisal: reflections from practice

Over the course of the 1990s, donor enthusiasm for participation came to be institutionalized in a variety of ways. One particular methodology—Participatory Rural Appraisal (PRA)—came to enjoy phenomenal popularity. New aid modalities may have shifted donor and lender concern away from the grassroots towards “policy dialogue.” But “civil society participation,” “social accountability,” and “empowerment”—some of the issues PRA grapples with—retain a place in the new aid discourse. PRA and its variants also continue to be used by government agencies, non-governmental, and community-based organizations in local-level assessment, planning, monitoring, and evaluation, as well as in national-level poverty assessments. It has sometimes been conflated, by donors and critics alike, with doing participatory development, and has elicited critiques that often go far beyond the bounds of the methodological. Yet these critics have tended to be academics with little experience as practitioners or facilitators. In this article, we draw on an action research project with PRA practitioners. We explore, through their critical reflections, some of the conundrums and contradictions faced by those who were active as PRA practitioners in the early 1990s. We suggest that the story of PRA’s success and of subsequent concerns about abuse and misuse by mainstream development institutions offers broader lessons with continued salience for development.