【作者】 邓小蕾；

【导师】 李民赞； Qin Zhang；

【作者基本信息】 中国农业大学 ， 农业电气化与自动化， 2014， 博士

【摘要】 果树在种植业中属高效农业。果树的生长状况与果树生长环境和果园管理密切相关。实时获取果树的生长环境信息及营养状况,可以为果园精细管理提供依据。无线传感器网络可以用来监测果树的生长环境信息及营养状况,避免了果园内布线的困扰。光谱技术可以对果树(苹果、冬枣等)的营养状况进行监测,避免了化学试验引起的人力和物力的浪费。高光谱图像可以同时获得空间信息和光谱信息,因而可以用来检测苹果采摘的机械损伤。因此,本论文开展了以下的主要研究内容：[1]基于无线传感器网络的果园信息采集系统基于无线传感器网络,开发了果园信息采集系统,主要用来采集空气温湿度、土壤温度、土壤水分等环境信息及表征作物长势的植被指数。系统由三部分组成：(1)集成ZigBee协调器、GPS模块、GPRS模块、RFID模块的PDA;(2)基于ZigBee的环境信息采集移动节点,用来采集空气温湿度、土壤温度、土壤水分等信息；(3)基于ZigBee的作物长势监测仪。PDA主要功能包括：采集GPS定位信息,采集RFID信息,与移动传感器节点间双向通信,将采集到的经纬度信息与果园信息(如土壤水分值)绑定后通过GPRS上传至远程上位机。环境信息采集移动节点的主要功能是：在刚加入到ZigBee网络以及加入到网络后接收到PDA的采集命令时,开启传感器电源并采集相应的传感器信息,然后向PDA里的ZigBee协调器发送数据,随后自动关闭传感器电源。作物长势监测仪的主要功能是实时采集8通道光强信息,以便后期组合处理成植被指数。硬件开发,主要包括PDA的集成与环境信息采集节点的硬件设计。软件开发,主要包括PDA的程序设计、自定义的通信协议、协调器与路由节点间的程序设计以及对实验室已有的作物长势监测仪的程序修改。系统进行了无线性能测试、温度传感器测试以及应用测试。试验表明,该系统能实时采集果园信息,保存并上传数据。[2]基于光谱技术的苹果与冬枣叶片营养元素检测应用光谱分析技术对苹果叶片的叶绿素含量和全氮含量,冬枣叶片的全氮含量进行了研究。讨论了不同预处理方法对模型的影响。应用全波段偏最小二乘回归(PLSR),特征波段偏最小二乘回归(PLSR)和支持向量机(SVM),建立了苹果叶片叶绿素含量和全氮含量的预测模型、冬枣叶片全氮含量的预测模型。讨论了样本集划分、异常样本判别与剔除、消除光谱噪声和其它因素干扰、小波与小波包降噪参数等预处理方法。讨论了5种样本集划分方法：随机抽样法(Random Sampling, RS)、含量梯度法(Content Gradient Method, CGM)、KS算法(Kennard-Stone Algorithm)、Duplex算法(Duplex Algorithm)、SPXY算法(Sample set Partitioning based on joint x-y distances Algorithm)。其中,SPXY划分方法和KS方法使PLSR模型预测精度最高。讨论了4种判别与剔除异常样本的方法：主成分得分图、正态分布箱型图、残差平方与杠杆值的关系图、基于蒙特卡洛的交叉验证法。除了主成分得分图法,其余3种方法剔除异常样品均能使PLSR-LOO-CV模型的RMSECV降低。讨论了反射率及其相应的吸光度在不同预处理方式下的PLSR模型。预处理方法有：归一化、Savitzky-Golay平滑(SG)、多元散射校正(MSC)、变量标准化校正(SNV)、去趋势(Detrending)、阶求导(1st derivative)。在这6种预处理方法中,无论是在KS样本集划分方法还是SPXY划分方法下,无论是剔除样本前还是剔除异常样本后光谱数据最佳预处理方式为归一化。讨论了小波与小波包降噪中的小波基函数、分解层数、默认阈值。综合考虑RMSECV与操作步骤,后续处理采用db4小波5层分解默认全局阈值。为了提高模型的预测精度,预处理可以采用(1)蒙特卡洛的交叉验证法或残差平方与杠杆值的关系图剔除异常样本；(2)归一化；(3)db4小波5层分解默认全局阈值去噪；(4)KS法或SPXY法划分样本集。测量了不同时间的苹果叶片叶绿素含量和全氮含量。采用SPXY样本集划分法,利用小波去噪、归一化、直接正交信号校正(DOSC)、连续投影算法(SPA)提取特征波长等预处理对苹果叶片叶绿素含量、全氮含量进行了全波段PLSR、特征波段SPA-PLSR、SVM建模。对于苹果叶片叶绿素含量的预测,线性方法DOSC-SPA-PLSR方法最佳。其在新梢生长期、花芽分化期、果实成熟期中,对新梢生长期的苹果叶片叶绿素含量预测效果最佳,模型的Rc、 RP、RMSEC、RMSEP、预测RPD分别为0.9980、0.9991、0.9704、0.7238、18.1841。而归一化后用全波段PLSR建模,其Rc、Rp、预测RPD分别为0.5345、0.6705、1.1665。对于苹果叶片全氮含量的预测,全波段PLSR模型和SPA-PLSR模型在新梢生长期与花芽分化期的建模与预测精度高于果实成熟期。因而,新梢生长期和花芽分化期为利用反射光谱检测叶片全氮含量的最佳时期。DOSC-SPA-PLSR方法最佳,其对新梢生长期的苹果叶片全氮含量预测效果最佳,Rc、Rp、预测RPD分别为0.9968、0.9969、11.1241。而归一化后用全波段PLSR建模,其Rc、RP、预测RPD分别为0.7420、0.5177、1.0433。对于冬枣叶片全氮含量的预测,DOSC-SPA-PLSR模型与原始冠层叶片反射率PLSR模型和DOSC-PLSR模型相比,减少了输入变量个数及建模潜变量LV个数,降低了模型的复杂度；同时,DOSC-SPA-PLSR模型的RMSEP比DOSC-PLSR模型的RMSEP小,预测精度有所提高。[3]基于高光谱图像的苹果采摘过程机械损伤检测苹果采摘过程机械损伤检测是开发采摘机器人的重要一环。采用机械手、压力传感器、NI数据采集卡等构建施力系统,用机械手夹持苹果并施加不同的压力,然后在不同的贮存时间拍摄高光谱图像。应用MATLAB对高光谱图像进行处理,提取了压力传感器与苹果接触区域和非接触区域的灰度平均值,并转化成光谱值。分析了红富士苹果偏红一侧与偏黄一侧的接触区域与非接触区域的光谱值随时间和压力的变化。计算了红富士苹果偏红一侧与偏黄一侧的接触区域与非接触区域的平均灰度值的标准偏差,从中提取波峰波谷值对应的多个单色图像进行比值、差值运算来实现对采摘过程中的苹果机械损伤的检测。该方法只适用于施加压力较小的苹果损伤检测,而在施加压力过大的情况下,则不能正确识别正常和受损区域。更多还原

【Abstract】 Fruit industry is efficient in planting industry. The growth of fruit trees is closely related to their growth environment and asscociated with orchard management. Real-time accessing to environmental information and the nutritional status of fruit trees could offer an approach to orchard management. Wireless sensor networks could be used to monitor environmental information and nutritional status of fruit trees, avoiding the wiring in the orchard. The nutritional content of fruit trees could be monitored by spectroscopy, avoiding the time comsuption of chemical diagnosis. The hyperspectral image could obtain the spatial and spectral information at the same time, thus could be used to detect the bruise on apples. The main content of the thesis are as follows:[1] Orchard information acquisition based on wireless sensor networksThe orchard information acquisition system was developed based on the wireless sensor networks to gather the air temperature and humidity, soil temperature, soil moisture, and vegetation index, etc. The system consisted of three parts:(1) PDA integrated with a ZigBee coordinator, GPS module, GPRS module, and RF1D module;(2) Environmental information acquisition nodes based on ZigBee, to gather the air temperature and humidity, soil temperature, soil moisture, etc.;(3) Crop growth status detection device based on ZigBee. The PDA was mainly used to collect the GPS information, RFID information, the ZigBee sensor node information, and then bind the orchard information (such as the soil moisture) and the GPS information to upload to the remote PC. The environmental information acquisition nodes were manily used to turn on the sensor power supply and gather the sensor information when receiving the gather command from the PDA. The crop growth status detection device was used to gather8-channel light intensity information, which could be processed to vegetation index later.Hardware design included the integration of PDA and the design of the environmental information acquisition nodes. Software design included the programming for PDA, the custom communication protocols, the programming for ZigBee coordinator and the Zigbee router, and the modification of the existing crop growth status detection device.The system test consisted of the test for the wireless performance, the test for the temperature sensor, and the application test. The application test showed that the system could get the realtime orchad information, save and then upload the data.[2] Nutrient content monitoring of apple leaves and jujube leaves using reflectanceThe chlorophyll content and nitrogen content of apple leaves and the nitrogen content of jujube leaf were measureed. The effects of different preprocessing methods on the models were discussed. Partial least squares regression (PLSR) through the full wavebands and characteristic wavebands PLSR and support vector machine (SVM) were built for a prediction of apple leaf chlorophyll content, apple leaf nitrogen content, and jujube leaves nitrogen content. Preprocessing methods such as sample set partitioning methods, outlier identification and elimination methods, noise elimination methods, and wavelet and wavelet packet noise denoising methods were discussed. Sample set partitioning methods such as random sampling method (RS), the content gradient method (CGM), Kennard-Stone Algorithm (KS), Duplex algorithm, sample set partitioning based on joint x-y distances algorithm (SPXY) were discussed. The SPXY partition method and KS method had the highest prediction accuracy. Identification and elimination methods such as principal component score map, boxplot, leverage versus squared residual plot, and the Monte Carlo cross-validation method were discussed. Except for the principal component score map, the other three methods could reduce the root mean square error of cross-validation (RMSECV). Preprocessing methods such as normalization, Savitzky-Golay smoothing (SG), multiplicative scatter correction (MSC), variable standardization correction (SNV), detrending, and first order derivative (1st derivative) were discussed. Normalization is the best one among all these six preprocessing methods Wavelet and wavelet packet de-noising were discussed, including the wavelet basis function, decomposition layers, and the default threshold. Considering the RMSECV and operation steps, db4wavelet5layer decomposition with default global threshold was chosen.Therefore, in order to improve the prediction accuracy, the following steps could be adopted:(1) Outlier detection by leverage versus squared residual plot or the Monte Carlo cross-validation method;(2) Normalization;(3) db4wavelet5layer decomposition with default global threshold;(4) KS or SPXY method to partition the sample set.The chlorophyll content and nitrogen content of apple leaf in different time were measured. SPXY algorithm was used to partition sample set. Wavelet denoising, normalization, direct orthogonal signal correction (DOSC), continuous projection algorithm (SPA) were used as the preprocessing methods for the full waveband PLSR, the selected wavebands PLSR, and the SVM modeling.For predicting apple leaf chlorophyll content, DOSC-SPA-PLSR modeling was the best modeling methods in its growing stages. In sprouting period, the Rc, Rp, RMSEC, RMSEP, RPD were0.9980,0.9991,0.9704,0.7238, and18.1841, respectively. During the same period, the Rc, RP, and RPD for normalized full waveband PLSR modeling were0.5345,0.6705, and1.1665, respectively.For predicting apple leaf nitrogen content, the full waveband PLSR modeling and SPA-PLSR modeling got better results in sprouting period and flower bud differentiation period. In sprouting period, Rc, Rp, RPD were0.9968,0.9969,11.1241, respectively. During the same period, Rc, RP, RPD for normalized full waveband PLSR modeling were0.7420,0.5177,1.0433, respectively.For predicting jujube leaf nitrogen content, DOSC-SPA-PLSR modeling reduced the input variables and the number of latent variables (LV). Meanwhile, DOSC-SPA-PLSR modeling reduced RMSEP, and improved the prediction accuracy.[3] Detection of apple bruises caused by picking with the manipulator based on hyperspectral imagesDifferent forces were applied on apples by a manipulator, and then hyperspectral images were taken after different storing days. As the Red Fuji apples were bi-color with red and yellow, the red side and yellow side of the apples were chosen to be the contacted area with the manipulator. Hyperspectral images (Hypercube data size600×1004×881) were processed in MATLAB. Average gray values of pixels selected from the contacted region and non-contacted region on both red and yellow side were extracted. They were then calibrated to obtain the reflectance respectively. The reflectance from hyperspectral images of the apples applied with different forces at the same time was compared. The reflectance from different storing time was also compared to observe changes over time after the forces applied by the manipulator. The standard deviation of the average gray value in contacted area and non-contacted area was calculated. The peak and valley wavelengths were seletcted. Then monochrome images at these wavelengths were selected to combine by ratio and difference operator to detect the apple bruices.更多还原