基于深度置信网络的农用地基准地价评估模型

该文针对现有基准地价评估模型主观性较强及精度不够等问题,提出了一种基于深度学习思想的农用地基准地价评估方法,构建了样本特征集合与地价标签集合的深层网络结构映射关系,并以广东省普宁市农用地基准地价评估为实例,验证了模型的可行性和有效性。结果如下：1）与神经网络、支持向量机这类浅层学习模型相比,深度置信网络模型其对地价的拟合精度要高出3.61%,3.12%;2）在训练样本只有300时,深度置信网络模型对地价的拟合误差为16.43%,比神经网络、支持向量机模型的拟合精度最少要高出6.76个百分点;3）深度置信网络模型对单个样本的运算时间及内存占用比神经网络、支持向量机模型都要高,但在评估精度都达到95%左右的情况下,深度置信网络模型所需训练样本较少,支持向量机训练时间为193 s,而深度置信网络模型耗时187 s,两者耗时基本持平;4）基于DBN模型对耕地评估单元的地价测算结果,将将普宁市耕地评估单元划分为5级,然后利用面积加权法求取对应的级别基准地价范围为21.34～26.23元/m2。上述试验结果表明该方法在对样本点地价的评估精度上要优于传统的浅层方法,并且该模型计算所得普宁市耕地基准地价与耕地质量在空间分布规律上保持一致。     

气力式秸秆深埋还田机输送装置设计与试验

针对合理耕层构建技术指标要求,设计了气力式秸秆深埋还田机输送装置。其主要结构参数为：输送管截面为0.2m×0.2m方形管;叶轮直径为0.55m,叶轮宽度为0.17m,进气口直径为0.26m,风机壳宽度为0.2m;螺旋轴直径为0.09m,螺旋叶片外径为0.25m,螺距为0.2m,螺旋叶片厚度为0.003m,螺旋外径与输送管内表面间隙为0.005m。通过玉米悬浮速度试验测得,长度为10cm玉米秸秆上、中、下部分悬浮速度分别为10.4、12.3、12.7m/s,平均值为11.9m/s,试验结果与仿真误差为7%。基于气固耦合理论,通过CFD-DEM气固耦合法对输送装置内的气固两相流模拟研究,表明弯角30°、转速为1800r/min时输送管道中,秸秆最小速度为5.21m/s,所对应的气流速度为17~27m/s,出口处玉米秸秆速度为6.06m/s,气流速度为2~27m/s,秸秆输送效果最佳。田间试验结果表明,气力输送装置性能参数最优组合为：风机转速1800r/min,秸秆覆盖量1.2kg/m2,叶片弯角30°。田间验证试验得深埋合格率为93.2%,有效提高了深埋质量。     

Application of modified Deep Belief Network in Forecasting Cotton Diseases and Insect Pests

[Objective] The occurrence and development of cotton diseases and insect pests are mainly related to environmental information. Because this environmental information is various, complex and unstable, the study on the prediction methods of cotton diseases and insect pests is a certain challenge. This study aims to establish a forecasting model for the timely and accurate prediction of cotton diseases and insect pests. [Method] A forecasting model of cotton diseases and insect pests is proposed based on environmental information and a modified Deep Belief Network (DBN) that is constructed by a three-layer restricted Boltzmann machine (RBM) and a supervised back-propagation (BP) network. In the method, the RBM is used to transform the original environmental information vectors into a new feature space related to the diseases and pests; the BP network is trained to classify and forecast the features generated by the last RBM layer and two rules of dynamic learning and comparison and dispersion are adopted to accelerate the training process of RBM. The proposed model was validated on a dataset of cotton bollworm, aphids, spider, cotton Verticillium wilt, and Fusarium wilt in a recent six-year period. [Result] Compared with the traditional prediction models of cotton diseases and insect pests, the proposed model can deeply explore the extensive correlation between the occurrence of cotton diseases and pests and environmental information. The results show that the proposed model has a higher accuracy compared with the classical predictive models, and the average forecasting accuracy is above 83%. [Conclusion] The proposed method is an effective crop disease and pest forecasting method that can provide a technical support for preventing cotton disease and insect pests.     

基于自动导航的小麦精准对行深施追肥机设计与试验

针对冬小麦返青期地表追施氮肥使氮素挥发导致肥料利用率低的问题,结合目前在小麦追肥过程中缺少深施氮肥作业装备的现状,进行了基于拖拉机自动导航技术实现精准对行深施氮肥的技术研究,设计了小麦精准对行精量深施追肥机。追肥机采用安装有自动导航系统的拖拉机牵引实现精准对行,以RTK-GNSS接收机测取的作业速度为基准,通过液压系统驱动排肥机构工作,双圆盘开沟器开沟深施,采用PID控制排肥轴转速与车辆行驶速度实现实时匹配,达到精量控制追肥量的目的。田间试验结果表明:设置目标追肥量为200 kg/hm2,车辆行驶速度为5 km/h时,追肥机能完成对行深施追肥作业,机具对行作业误差在±6 cm以内,追肥量偏差小于9%,可满足实际生产需求;对照撒肥机表层撒肥作业,每公顷减施氮肥25 kg左右,小麦每公顷增产486.5 kg左右。     

基于深度信念网络的猪咳嗽声识别

为了在生猪养殖产生呼吸道疾病的初期,通过监测猪咳嗽声进行疾病预警,提出了基于深度信念网络(DBN)对猪咳嗽声进行识别的方法。以长白猪咳嗽、打喷嚏、吃食、尖叫、哼哼、甩耳朵等声音为研究对象,利用基于多窗谱的心理声学语音增强算法和单参数双门限端点检测对猪声音进行预处理,实现猪声音信号的去噪和有效信号检测。基于时间规整算法提取300维短时能量和720维梅尔频率倒谱系数(MFCC)组合成1020维特征参数,将该组合特征参数作为DBN学习和识别数据集,选定3隐层神经元个数分别为42、17和7,构建网络结构为1020-42-17-7-2的5层深度信念网络猪咳嗽声识别模型。通过5折交叉实验验证,基于DBN的猪咳嗽声识别率和总识别率均达到90%以上,误识别率不超过8.07%,最优组猪咳嗽声识别率达到94.12%,误识别率为7.45%,总识别率达到93.21%。进一步基于主成分分析法(PCA)提取1020维特征参数98.01%主成分得到479维特征参数,通过5折交叉实验验证,猪咳嗽声识别率和总识别率相对降维前均有所提高,误识别率有所降低,最优组猪咳嗽声识别率达到95.80%,误识别率为6.83%,总识别率达到94.29%,实验结果表明所建模型是有效可行的。     

Peanut macronutrient absorptions characteristics in response to soil compaction stress in typical brown soils under various tillage systems

Soil tillage, a major agricultural management, could effectively alter soil structure and plant growth, particularly under groundnut plantations. To understand effects of different tillage measures on nitrogen(N), phosphorus(P) and potassium(K) absorptions and use efficiencies for peanut (Arachis hypogaea L.), four tillage treatments: no tillage (NT), deep loosing (DL), deep plow (DP), and shallow plow (SP), were examined for two growing years at three typical peanut-producing sites of Qishan, Wangcheng, and Xiadian in Shandong, China. Results showed that average soil bulk density under DL, DP, and SP at the three sites was decreased by 7.1–19.5% compared with NT treatment for the 2 years. Significantly higher average total N accumulations in underground peanut part patterned as DP (163 kg/ha) > SP (149 kg/ha) > DL (144 kg/ha) > NT (117 kg/ha), while total N in aboveground peanut part was 8.7–22.1% higher under DP than other treatments. Absorptions of N, P, and K in underground parts were extremely significantly contributed to high peanut yields (P < 0.01), whereas increase of N and P absorptions in aboveground parts did not promote peanut yields. Soil bulk density was significantly negatively correlated with plant macronutrient amounts in underground peanut parts and peanut yields (P < 0.01). Moreover, N:P, N:K, and P:K ratios were similar between NT and noncompaction stress treatments of DL, DP, and SP. These results indicate that DP is a rational tillage practice for promoting nutrient uptake amount, efficiency, and peanut yields by alleviating soil compaction stress in peanut-producing fields.     

基于卷积神经网络的中国北方冬小麦遥感估产

针对传统的农作物估产方法过度依赖人工经验,以及实地采样成本高等问题。该研究使用MODIS数据构建了基于卷积神经网络(convolutional neural network,CNN)的冬小麦估产模型。对2006-2016年中国北方冬小麦核心区的60个地级市进行模型训练,鲁棒性检验以及估产误差空间特征分析。结果表明:1)估产模型在训练集和验证集的均方根误差(root mean squareerror,RMSE)分别为183.82kg/hm2、689.72 kg/hm2,决定系数(R2)分别为0.98、0.71。2)以同样的神经网络结构对2006-2016年估产样本分别作为验证集,训练11个独立模型的RMSE平均值是772.03 kg/hm2,证明算法具有较高的鲁棒性。3)2007、2012和2016年不同省份的估产结果表明,模型对北方冬小麦区的平原区估产精度较高,尤其是河北和山东2省(RMSE为500 kg/hm2)。该文构建的估产模型可以实现冬小麦单产的复杂拟合,可以应用于较大尺度(范围)冬小麦产量预报。     

多通道深度可分离卷积模型实时识别复杂背景下甜菜与杂草

针对实际复杂田间环境下杂草与作物识别精度低和实时性差的问题,为减少弱光环境对分割识别效果的影响,实现甜菜与杂草的实时精确分割识别,该文首先将可见光图像进行对比度增强,再将近红外与可见光图像融合为4通道图像;将深度可分离卷积以及残差块构成分割识别模型的卷积层,减少模型参数量及计算量,构建编码与解码结构并融合底层特征,细化分割边界。以分割识别精度、参数量以及运行效率为评价指标,通过设置不同宽度系数以及输入图像分辨率选出最优模型。试验结果表明:本文模型的平均交并比达到87.58%,平均像素准确率为99.19%,帧频可达42.064帧/s,参数量仅为525 763,具有较高分割识别精度和较好实时性。该方法有效实现了甜菜与杂草的精确实时识别,可为后续机器人精确除草提供理论参考。     

人字形板式换热器流道传热特性及参数优化

为实现田间条件下快速、准确识别棉花与杂草,该文以自然光照下田间棉花与杂草为研究对象,采用垂直向下拍摄的方式获取棉花杂草视频,按1帧/s的速率从视频中提取图像,在人工去除冗余度过多的图片后,建立1 000幅图片的数据集。对比了Faster R-CNN和YOLOv3 2种典型卷积神经网络,将Faster R-CNN卷积神经网络的深度学习模型引入到棉花杂草图像识别中,并提出一种结构优化的方法,使之适用于复杂背景下的棉田杂草识别。该文选用残差卷积网络提取图像特征,Max-pooling 为下采样方法,RPN网络中引入特征金字塔网络生成目标候选框,对卷积神经网络结构进行优化。在使用700幅图片进行训练后,通过200 幅田间棉花杂草图像识别测试,结果表明：该方法的平均目标识别准确率达95.5%,识别单幅图像的平均耗时为1.51 s,采用GPU 硬件加速后识别单幅图像的平均耗时缩短为0.09 s。优化后的Faster R-CNN卷积神经网络相对于YOLOv3平均正确率MAP高0.3以上。特别是对于小目标对象,其平均正确率之差接近0.6。所提方法对复杂背景下棉花杂草有较好的检测效果,可为精确除草的发展提供参考。     

基于高光谱和深度迁移学习的柑橘叶片钾含量反演

针对传统柑橘叶片钾含量检测方法耗时费力、操作繁琐且损伤叶片等弊端,引入高光谱信息探索柑橘叶片钾含量快速无损检测与预测模型,选用ASD Field Spec 3光谱仪采集柑橘4个重要物候期(萌芽期、稳果期、壮果促梢期和采果期)的叶片反射光谱,同步采用火焰光度法测定叶片的钾含量;先用正交试验确定小波去噪的最佳去噪参数组合,再进行不同光谱形式变换,对不同物候期光谱进行基于堆栈稀疏编码机-深度学习网络(Stacked sparse autoencoder-deep learning networks,SSAE-DLNs)的特征提取迁移和融合多种特征,对比支持向量机回归、偏最小二乘法回归、广义神经网络、逐步多元线性回归等多种诊断模型,结果表明,模型SSAE-DLNs基于一阶微分光谱特征建立全生长期钾含量预测模型的性能最优,其校正集和验证集决定系数分别为0. 898 8、0. 877 1,均方根误差分别为0. 544 3、0. 552 8。试验表明,深度迁移学习网络可对柑橘叶片钾含量进行精确预测,为高光谱检测技术用于柑橘树长势监测和营养诊断提供了参考。