基于语义分割的非结构化田间道路场景识别

环境信息感知是智能农业装备系统自主导航作业的关键技术之一。农业田间道路复杂多变,快速准确地识别可通行区域,辨析障碍物类别,可为农业装备系统高效安全地进行路径规划和决策控制提供依据。该研究以非结构化农业田间道路场景为研究对象,根据环境对象动、静态属性进行类别划分,提出一种基于通道注意力结合多尺度特征融合的轻量化语义分割模型。首先采用Mobilenet V2轻量卷积神经网络提取图像特征,将混合扩张卷积融入特征提取网络最后2个阶段,在保证特征图分辨率的基础上增加感受野并保持信息的连续性与完整性;然后引入通道注意力模块对特征提取网络各阶段特征通道依据重要程度重新标定;最后通过空间金字塔池化模块将多尺度池化特征进行融合,获取更加有效的全局场景上下文信息,增强对复杂道路场景识别的准确性。语义分割试验表明,不同道路环境下本文模型可以对场景对象进行有效识别解析,像素准确率和平均像素准确率分别为94.85%、90.38%,具有准确率高、鲁棒性强的特点。基于相同测试集将本文模型与FCN-8S、SegNet、DeeplabV3+、BiseNet模型进行对比试验,本文模型的平均区域重合度为85.51%,检测速度达到8.19帧/s,参数数量为,相比于其他模型具有准确性高、推理速度快、参数量小等优点,能够较好地实现精度与速度的均衡。研究成果可为智能农业装备在非结构化道路环境下安全可靠运行提供技术参考。     

丘陵山区田间道路自主行驶转运车及其视觉导航系统研制

自然条件的限制使得丘陵山区农产品和物资的田间转运难以实现高安全性的机械化作业。为此,该文研制了一种在丘陵山区田间道路上自主行驶的转运车及其视觉导航系统。该系统采用RTK-GNSS(real-timekinematic-global navigationsatellitesystem,实时动态-全球卫星导航系统)进行路网信息采集、实时定位和路径规划,利用机器视觉进行田间道路识别并提取路径跟踪线;田间道路非路口区域由机器视觉系统进行导航,路口区域采用RTK-GNSS实时定位进行导航。全局路径规划中对A*算法估价函数进行改进,将路口节点处的道路曲率及道路起伏信息引入代价函数。图像处理中强化道路上的阴影处理和信息融合,实现道路与背景的准确分割;然后将道路区域分块求取形心点,拟合后生成道路的虚拟中线作为局部路径的导航线。路径规划仿真表明,改进的A*算法能融合丘陵山地道路起伏变化的特征,规划的路径更合理。转运车自主行驶测试表明,在直线路径、多曲率复杂路径以及地形起伏路径3种工况下,自主行驶轨迹与实际道路中线的平均偏差分别为0.031、0.069和0.092 m,最大偏差分别为0.133、0.195和0.212 m;转运车沿道路中线自主行驶的平均相对误差分别为5.16%、11.5%和15.3%,满足田间道路转运车自主行驶的安全要求。     

基于多学科技术融合的智能农机控制平台研究综述

农业机械的自动化和智能化包含内容广泛,有农机定位与导航,动态路径规划,机器视觉和远程监控等,牵涉到大量的工程技术学科,包括导航、图像、模型与策略、执行器以及数据链等。农机定位与导航一般采用基于农机运动学模型结合GPS(global positioning system)/IMU(inertial measurement unit)组合导航信息,在导航路径规划算法指引下实现农机轨迹跟踪的方法。建立的农机运动学模型精度,GPS数据的连续性以及惯导器件误差系数漂移等因素都会影响该方法的有效性。路径跟踪通常采用各种现代控制理论与方法,而面对复杂的田间作业环境变化,农机的自主避障以及动态路径规划能力也会影响轨迹跟踪精度。机器视觉的稳定性和目标特征信息分离度影响着农机环境感知能力,目前目标识别主要采用hough变换,hough变换的全局检测特性决定了该算法运算量较大,需要探究改进特征提取算法。远程监控农机作业是智能农机发展的一个方向,构建无线导航,控制和视频数据传输网络有助于提高农机的智能化水平,可以采用分布式哈希表(distributed hash table)来研究网络覆盖和互联技术。该文融合多个学科,从高精度定位与导航技术、复杂环境及工况下农机运动精确自主控制技术、稳定清晰的机器视觉感知技术和基于4G网络和新一代物联网的高覆盖数据传输技术几个方面,论述了智能农机在光机电液多个学科领域内的研究现状,并指出采用北斗地基增强网络和网络RTK(real-time kinematic)技术、惯导定位误差精确建模与补偿、环境感知与自主避障、立体结构自组网技术以及多机协作是现代农业机械的发展方向。以期为现代化智能农业机械的设计提供参考。     

农业车辆双目视觉障碍物感知系统设计与试验

为保证智能化农业机械在自主导航过程中的安全可靠性,该研究将嵌入式AI计算机Jetson TX2作为运算核心,设计一套基于双目视觉的农业机械障碍物感知系统。使用深度卷积神经网络对作业环境中的障碍物进行识别,并提出一种基于改进YOLOv3的深度估计方法。将双目相机抓取的左右图像分别输入至改进的YOLOv3模型中进行障碍物检测,并将输出的目标检测框信息进行目标匹配和视差计算,完成对障碍物的识别、定位和深度估计。试验结果表明,该系统能够对障碍物进行准确识别,平均准确率和召回率分别达到89.54%和90.18%;改进YOLOv3模型的深度估计误差均值、误差比均值较原始YOLOv3模型分别降低25.69%、25.65%,比Hog+SVM方法分别降低41.9%、41.73%;动态场景下系统对障碍物深度估计的平均误差比为4.66%,平均耗时0.573s,系统在深度预警时能够及时开启电控液压转向模块进行安全避障。研究结果可为农业机械的自主导航提供有效的环境感知依据。     

基于改进FCN-8s的灵武长枣图像分割方法

针对不同成熟度灵武长枣图像中多尺度长枣目标的分割问题,该研究提出了一种基于改进FCN-8s的灵武长枣图像分割方法。首先,建立包含不同成熟度灵武长枣图像的数据集。其次,提出一个多尺度特征提取模块,该模块以3×3卷积为主分支,增加1×1卷积和5×5深度可分离卷积作为辅助分支,以提取灵武长枣图像中的多尺度特征。然后,用多尺度特征提取模块替换FCN-8s中的3×3卷积,又对FCN-8s网络结构做了一定的改进,得到了改进FCN-8s。最后,在灵武长枣数据集上进行试验,结果表明,改进FCN-8s的枣类交并比、平均交并比、像素准确率、召回率和像素准确率、召回率的调和平均值F_1分数分别达到了93.50%、96.41%、98.44%、97.86%和98.15%,比原FCN-8s的各项评价指标分别高出了11.31、6.20、1.51、5.21和3.14个百分点。网络参数量为5.37×106,分割速度为16.20帧/s。改进FCN-8s满足了灵武长枣智能化采摘机器人对视觉识别系统的要求,为实现灵武长枣的智能化采摘提供了技术支持。     

基于通道特征金字塔的田间葡萄实时语义分割方法

复杂环境下葡萄的快速检测识别是智能采摘的关键步骤,为解决目前葡萄识别精度低和实时性差的问题,该研究提出一种轻量级葡萄实时语义分割模型（Grape Real-time Semantic Segmentation Model,GRSM）。首先,利用通道特征金字塔（Channel-wise Feature Pyramid,CFP）模块进行特征提取,该模块通过1?3和3?1空洞卷积的跳跃连接,在减少模型参数量的同时提取葡萄图像的多尺度特征和上下文信息;然后,采用池化卷积融合结构完成下采样,增加可训练参数以减少信息损失;最后,利用跳跃连接融合多种特征恢复图像细节。试验结果表明：该研究所提出的模型在田间葡萄测试集上达到了78.8%的平均交并比,平均像素准确率为90.3%,处理速度达到68.56帧/s,网络结构大小仅为4.88MB。该模型具有较高分割识别精度和较好实时性,能满足葡萄采摘机器人对视觉识别系统的要求,为葡萄的智能化采摘提供了理论基础。     

基于TMU-Net网络的苹果果心分割方法

针对苹果内在品质检测过程中传统测量果心大小方法效率低、准确性差等问题,该研究提出一种基于TMU-Net网络自动分割果心的方法,将Transformer编码器融入U-Net网络结构中,构建改进U型卷积网络TMU-Net模型。模型由特征提取模块、特征处理模块、解码器、特征拼接模块组成,以VGG-16前13层作为主干特征提取网络,在跳跃连接中叠加多重残差空洞卷积（Multiple Residual Dilated Convolution,MRDC）模块,增大感受野的同时增强了模型对底层特征提取能力。采用数据增强技术对果心数据集扩充后,利用迁移学习方法冻结特定的网络层,对TMU-Net模型进行训练。试验结果表明：引入迁移学习并使用最佳训练方式使模型分割精确率提高了22.48个百分点;TMU-Net网络模型在果心分割任务中实现了96.72%的精确率,与U-Net、PSPNet、DeeplabV3+网络对比,精确率分别提升了14.28、9.98、7.15个百分点。该方法能够精准、有效地实现果心分割,可为实现苹果内在品质智能检测提供参考。     

高地面仿形性动力底盘的设计与试验

针对一般农业机械动力底盘在丘陵山区行驶、田间作业时,由于地形适应性差以及附着力不足导致车轮打滑,影响整机作业效果等现象,采用地面仿形原理,设计一种全时8轮驱动、5自由度仿形的山地农业机械动力底盘,并对底盘主要结构参数进行了分析计算。试验证明：底盘样机具有良好的地形适应性和通过性,且转向灵活,能够保持良好的附着力,满足在崎岖不平的道路行驶及田间作业要求。     

一种基于GPS和GIS农业装备田间位置的监控系统

对田间车辆的实时监控和导航，是实施精细农业变量作业技术的基础。该文结合中国国情和旱作农业的实际，将GPS、GIS技术相结合，采用MapInfo公司生产的MapX4.5控件内嵌可视化编程语言Visual Basic 6.0，研制开发了一种基于GPS和GIS的田间农业装备实时监控和信息管理系统。系统完成了地图常用功能和各种GIS工具模块的设计,包括串口通讯、地图编辑、网络通讯3个子系统，共17个模块，本系统采用控件技术设计了串口和网络的数据传输和传输数据的保存；以旱作农业机械为例，建立了动态田间农业机械装备数     

基于无人机图像的多尺度感知麦穗计数方法

小麦是世界上重要的粮食作物,其产量的及时、准确预估对世界粮食安全至关重要,小麦穗数是估产的重要数据,因此该研究通过构建普适麦穗计数网络（Wheat Ear Counting Network,WECnet）对灌浆期小麦进行精准的计数与密度预估。选用多个国家不同品种的麦穗图像进行训练,并且对数据集进行增强,以保证麦穗多样性。在原始人群计数网络CSRnet基础上,针对小麦图像特点构建WECnet网络。在网络前端,通过使用VGG19的前12层进行特征提取,同时与上下文语义特征进行融合,充分提取麦穗的特征信息。后端网络使用不同空洞率的卷积加大感受野,输出高质量的密度图。为了验证模型的可迁移性与普适性,该研究通过基于全球小麦数据集训练好的模型对无人机实拍的麦田图像进行计数。试验结果表明：在全球小麦数据集上,WECnet训练模型的决定系数、均方根误差（Root Mean Square Error,RMSE）与平均绝对误差（Mean Absolute Error,MAE）分别达到了0.95、6.1、4.78。在无人机拍摄图像计数中,决定系数达到0.886,整体错误率仅为0.23%,平均单幅小麦图像计数时间为32 ms,计数速度与精度均表现优异。结果表明,普适田间小麦计数模型WECnet可以对无人机获取图像中小麦的准确计数及密度预估提供数据参考。     

低氮和干旱胁迫对富士和秦冠生长及氮素利用的影响

【目的】以富士（Fuji）、 秦冠（Qinguan）嫁接在平邑甜茶（Malus hupehensis Rehd.）上的当年生盆栽苗为试验材料,采用砂培方法,研究了缺氮胁迫和干旱对富士和秦冠生长情况、 光合参数、 植株各部位氮磷钾含量及氮素利用效率的影响,分析比较了低氮干旱条件下富士和秦冠生长及氮素利用的差异,以期为果树生产高效肥水利用提供理论指导。【方法】试验共设四个处理: 正常氮正常水（ZZ）、 低氮正常水（DZ）、 正常氮干旱（ZG）、 低氮干旱（DG）。氮素和水分均设置两个水平,分别为正常氮（6 mmol/L NO-3-N）、 低氮（0.3 mmol/LNO-3-N）、 正常供水（保持盆中砂子相对含水量为饱和含水量的80%~85%）、 干旱处理（保持盆中砂子相对含水量为饱和含水量的60%~65%）。【结果】富士和秦冠的生物量（茎和叶）、 株高茎粗等生长指标以及光合速率、 气孔导度、 蒸腾速率均为正常氮正常水（ZZ）＞低氮正常水（DZ）＞正常氮干旱（ZG）＞低氮干旱（DG）,并且相对应处理下秦冠的以上指标均高于富士;正常供水下,缺氮处理使富士、 秦冠的根冠比比正常氮处理均有所增加,富士提高了2.05%,秦冠提高了22.40%。富士和秦冠的氮、 磷、 钾含量均表现出正常氮正常水（ZZ）＞低氮正常水（DZ）＞正常氮干旱（ZG）＞低氮干旱（DG）; 氮、 钾元素含量在植株各部位的分布顺序依次是叶＞根＞茎,磷元素则是根＞叶＞茎;光合氮素利用效率（PNUE）和氮素利用效率表现为秦冠处理之间差异极显著,富士处理之间差异不显著;秦冠的PNUE和NUE明显高于富士,在低氮正常水（DZ）处理下,秦冠氮肥利用率比富士高42.07%,在低氮干旱（DG）处理下高64.14%;低氮胁迫下富士和秦冠的NUE显著提高,并且秦冠提高的幅度高于富士。【结论】施用氮肥能够显著提高富士与秦冠的干物质量,同等水肥条件下,秦冠生长优于富士;水分亏缺会减少叶片对氮的吸收,干旱条件下适度增施氮肥,可提高果树的抗旱能力;低氮干旱胁迫下秦冠的生长指标、 光合指标及氮素利用效率指标均优于富士,表现出较强的抗低氮干旱胁迫的能力。     

Pasture Vegetation Elemental Analysis by Laser-Induced Breakdown Spectroscopy

Laser-induced breakdown spectroscopy (LIBS) is a new technique for the analysis of plant material. This study investigates the application of LIBS to pasture-based plant samples. The LIBS measurements were obtained from pelletized pasture samples (100 samples) that had been also analyzed by inductively coupled plasma–optical emission spectroscopy (ICP-OES) following microwave digestion for calibration and comparison purposes. Comparisons for elements sodium (Na), potassium (K), magnesium (Mg), calcium (Ca), manganese (Mn), iron (Fe), copper (Cu), zinc (Zn), boron (B), phosphorus (P), and sulfur (S) showed that LIBS could be used for almost all the standard profile total elements with concentrations down to low mg/kg levels (observed error of Na: 0.024 percent, K: 0.18 percent, Mg: 0.016 percent, Ca: 0.073 percent, P: 0.017 percent, Mn: 31 mg/kg, Fe: 150 mg/kg, Zn: 6.6 mg/kg, and B: 1.1 mg/kg). Elemental analysis at less than mg/kg levels was not possible using LIBS. The elements S and Cu were particularly difficult to analyze with reliability using LIBS at the concentration levels found in the plant samples. Replacing microwave digestion and subsequent ICP analysis with a direct analysis of dried plant samples using LIBS has the potential to improve the productivity and reduce the cost of testing.     

Faculty positions:Center for Agricultural Resources Research,Chinese Academy of Sciences

正The Center for Agricultural Resources Research(CARR),the Institute of Genetics and Developmental Biology(IGDB),Chinese Academy of Sciences,invites applicants for several research group leader positions.CARR is one of the research organizations in Chinese Academy of Sciences(CAS).We seek nominations and applications from individuals who have expertise and a record of accomplishment in research areas related to ecology,agro-hydrology,     

Biodiversity and ecogeography of wild Lactuca spp. in some European countries

Staff members of the Department of Botany of Palacký University in Olomouc and Gene Bank Department – Workplace Olomouc, Research Institute of Crop Production in Prague, Czech Republic, conducted an expedition in seven European countries (Austria, Czech Republic, France, Germany, Italy, the Netherlands, Switzerland) in August/September 1999 to collect wild Lactuca spp. germplasm and study its geographic distribution, ecology and biodiversity. During the mission, more than 600 locations were visited resulting in the collection of 602 seed samples (accessions) of wild Lactuca species and 13 seed samples of related genera (Chondrilla and Mycelis). Lactuca serriola f. serriola, L. serriola f. integrifolia, L. saligna and L. viminea subsp. chondrilliflora were prevalent in southern Europe (Italy, France), however, only L. serriola was common in central and western Europe (Austria, Czech Republic, Germany, Netherlands, Switzerland). The greatest diversity of Lactuca species was found in France, where also the most seed samples (165) were collected. The most characteristic habitats with a high density of Lactuca spp. populations were observed along roads and highways, grassy ditches, ruderal communities, and dust-heaps. Natural infections by powdery mildew (Erysiphe cichoracearum) and downy mildew (Bremia lactucae) on some wild Lactuca spp. were observed. Recent observations concerning the geographic distribution, population structure, habitats, and natural occurrence of diseases of Lactuca spp. are discussed. This assemblage of genetic resources of Lactuca spp. can serve as the basis of future studies of species diversification, spatial population structure, plant microevolution, domestication processes, and genetic variability of host-parasite interactions.     

Macroinvertebrates in North American tallgrass prairie soils: effects of fire, mowing, and fertilization on density and biomass

The responses of tallgrass prairie plant communities and ecosystem processes to fire and grazing are well characterized. However, responses of invertebrate consumer groups, and particularly soil-dwelling organisms, to these disturbances are not well known. At Konza Prairie Biological Station, we sampled soil macroinvertebrates in 1994 and 1999 as part of a long-term experiment designed to examine the effects and interactions of annual fire, mowing, and fertilization (N and P) on prairie soil communities and processes. For nearly all taxa, in both years, responses were characterized by significant treatment interactions, but some general patterns were evident. Introduced European earthworms (Aporrectodea spp. and Octolasion spp.) were most abundant in plots where fire was excluded, and the proportion of the total earthworm community consisting of introduced earthworms was greater in unburned, unmowed, and fertilized plots. Nymphs of two Cicada genera were collected (Cicadetta spp. and Tibicen spp.). Cicadetta nymphs were more abundant in burned plots, but mowing reduced their abundance. Tibicen nymphs were collected almost exclusively from unburned plots. Treatment effects on herbivorous beetle larvae (Scarabaeidae, Elateridae, and Curculionidae) were variable, but nutrient additions (N or P) usually resulted in greater densities, whereas mowing usually resulted in lower densities. Our results suggest that departures from historical disturbance regimes (i.e. frequent fire and grazing) may render soils more susceptible to increased numbers of European earthworms, and that interactions between fire, aboveground biomass removal, and vegetation responses affect the structure and composition of invertebrate communities in tallgrass prairie soils.     

Properties of anaerobically digested and composted municipal solid waste assessed by linking soil mesofauna dynamics and nitrogen modelling

We have studied the effect of anaerobically digested (ADMSW) and composted municipal solid waste (CMSW) on mineralization and foodweb dynamics to verify the hypothesis that ADMSW would immobilize N right after addition to soil in contrast to addition with CMSW. Another hypothesis was that the mesofauna (enchytraeids and microarthropods) would stimulate N release from the decomposer community. We measured excretion of -N and urea-N from the mesofauna and hypothesized that enchytraeids would release urea. ADMSW and CMSW were amended to pots with sandy loam and barley. The pots were divided into treatments with or without mesofauna. Mesofauna, plant N and biomass, soil N and ergosterol (fungal biomass) were measured over a 113-day period of four equidistant samplings. Soil respiration, N mineralization and N release by the mesofauna were modelled from concurrent studies. ADMSW- and CMSW-treated soils initially (<20 days) immobilized N. The amendments did not increase plant growth substantially, and this was probably due to N-limitation in the early stages of plant growth. Enchytraeid abundance was about three times higher in ADMSW- than CMSW-treated soils, indicating that ADMSW contained more labile compounds, bacteria, and microfauna. The mesofauna did not affect N-content, but the cumulated -N excreted by the mesofauna was estimated to be substantial and about one-fifth of total plant N in ADMSW. An explanation for this discrepancy might be that in the absence of the mesofauna, other members of the detrivore and microbivore community performed the mesofauna’s function. Neither enchytraeids nor microarthropods excreted urea.     

Faculty positions: Center for Agricultural Resources Research,Chinese Academy of Sciences

正The Center for Agricultural Resources Research(CARR),the Institute of Genetics and Developmental Biology(IGDB),Chinese Academy of Sciences,invites applicants for several research group leader positions.CARR is one of the research organizations in Chinese Academy of Sciences(CAS).We seek nominations and applications from individuals who have expertise and a record of accomplishment in research areas related to ecology,agro-hydrology,     

Faculty positions: Center for Agricultural Resources Research,Chinese Academy of Sciences

正The Center for Agricultural Resources Research(CARR),the Institute of Genetics and Developmental Biology(IGDB),Chinese Academy of Sciences,invites applicants for several research group leader positions.CARR is one of the research organizations in Chinese Academy of Sciences(CAS).We seek nominations and applications from individuals who have expertise and a record of accomplishment in research areas related to ecology,agro-hydrology,     

Variation in quality parameters between and within 14 Nordic tree fruit and berry species

Abstract

A 3-year study was carried out to investigate quality parameters in 14 tree fruit and berry species grown in southern Norway. The species were blueberry, apple, aronia, sour cherry, sweet cherry, red raspberry, strawberry, blackcurrant, gooseberry, red currant and elderberry, harvested along with wild bilberry, cloudberry and lingonberry. Significant differences between species were identified for all quality parameters. The coefficient of variation between species was lowest for pH (12.5%), dry matter (18.9%) and soluble solids (25.3%), followed by titratable acids (59.3%), total phenolics (83.8%), antioxidant capacity FRAP (85.7%) and antiradical power by the DPPH-assay (97.8%), total monomeric anthocyanins (132%) and ascorbic acid (137%). Average coefficient of variation within species were lower and ranged from 4 (pH) to 62% (ascorbic acid). Only the FRAP values were significantly affected by harvesting year with lower levels in 2004 than in 2005 and 2006. There were significant interactions between species and harvesting year for dry matter, soluble solids, pH, ascorbic acid and FRAP. The results indicate generic ranges in composition within species independent upon growing location and climate, and the composition of the tree fruits and berries is not likely to deviate from these ranges. It is concluded that desirable composition of tree fruits and berries and their products should primarily be achieved by selection among species rather than searching fors broadened variation within individual species.     

Potassium Fixation and Release Characteristics in Rhizosphere and Nonrhizosphere Soils for a Rapeseed–Rice Cropping Sequence

Potassium (K) fixation and release in soil are important factors in the long-term sustainability of a cropping system. Changes in K concentration and characteristics of K fixation and release in rhizosphere and nonrhizosphere soils in the rapeseed (Brassica napus L.)–rice (Oryza sativa L.) rotation were investigated using a rhizobox system. The concentrations of different forms of K in both rhizosphere and nonrhizosphere soils decreased with plants compared to without plants, regardless of K fertilizer application. Potassium uptake by crops mainly came from the rhizosphere soil. In the treatment without K fertilizer (–K), the main form of K supplied by the soil to the crops was 1.0 mol L^?1 nitric acid (HNO₃) nonextractable K, followed by nonexchangeable K, and then exchangeable K. In the treatment with K fertilizer (+K), the main K forms supplied by the soil to the crops were exchangeable K and nonexchangeable K. The amount and rate of K fixation after one cycle of the rapeseed–rice rotation was greater in rhizosphere soil than in nonrhizosphere soil. The amount and rate of K fixation of soil in the +K treatment were significantly less than in the –K treatment. The cumulative amounts of K released with 1.0 mol L^?1 ammonium acetate (NH₄OAc) and 1.0 mol L^?1 HNO₃ extraction increased with the increasing numbers of extractions, but the K-releasing power of soil by successive extraction decreased gradually and finally became almost constant. The release of K was less in rhizosphere soil than in nonrhizosphere soil. The release of K in the +K treatment was similar to that in the –K treatment in rhizosphere soil, but the K release in nonrhizosphere soil was greater with the +K than the –K treatment. Overall, the information obtained in this study will be helpful in formulating more precise K fertilizer recommendations for certain soils.