基于多源多时相遥感数据水稻长势参数提取与应用

发布时间：2018-07-06 18:43

本文选题：水稻 + 高光谱　；参考：《华中农业大学》2017年硕士论文

【摘要】：水稻是主要粮食作物之一,其生长状况的监测对水稻生长管理、产量预测、灾害防治等方面起到重要作用。而叶面积指数、色素含量等生理生化参数是水稻长势监测的重要指标。凭借高光谱技术的波段连续性强、波谱分辨率高、光谱信息丰富等优势,能够实时、快速、高效、无损地获取水稻长势、营养状况及其变化状况,为精准农业的信息化管理提供技术支持和理论依据。本文以不同氮肥水平、不同生育期的水稻为研究对象,建立水稻生理生化参数的冠层反射光谱反演模型,基于环境卫星(HJ-1A)影像数据,获得水稻关键生长期的叶面积指数(Leaf area index,LAI)的空间分布,实现水稻生长状况的大范围观测。围绕上述内容,开展研究,取得了如下主要结果:1.基于水稻冠层反射光谱数据得到15个植被指数,利用留一法交叉验证进行5种传统回归模型分析(线性函数、指数函数、对数函数、幂函数、多项式函数),建立了不同生育期的水稻LAI高光谱指数估计模型,获取了不同生育期的优选植被指数,采用噪声等效误差(Noise Equivalent,NE)对植被指数反演LAI进行了敏感性分析,结果显示,分蘖期归一化植被指数(normalized difference vegetation index,NDVI)、新型植被指数(new vegetation index,NVI)对LAI变化敏感,且估计精度高;拔节期绿色归一化植被指数(green normalized difference vegetation index,GNDVI)、比值植被指数(ratio vegetation index,RVI-3)、改进的简单比值指数(modified simple ratio index,MSR)具有高敏感性和估计精度;水稻生长后期GNDVI、修正归一化差异植被指数(modified normalized difference vegetation index,mNDVI)、MSR比其他指数更适用于LAI估计。基于植被指数构建的全生育期水稻LAI传统回归模型精度过低,难以用统一植被指数来估算整个生育期水稻LAI,利用偏最小二乘回归建模方法,建模集和验证集的决定系数R2分别可达到0.87和0.81,RMSEC为0.612,RMSEP为0.856,RPD大于2,能够较为精确地估算全生育期水稻LAI。2.基于连续统去除处理的水稻冠层高光谱数据(400~750 nm),选取了波段深度(band depth,BD)、波段深度比(band depth ratio,BDR)、归一化波段深度(normalized band depth index,NBDI)和归一化面积波段指数(band depth normalized to band area,BNA)4种波段指数。在此基础上进行主成分分析(principal component analysis,PCA)实现光谱降维,然后运用反向传播(back propagation,BP)神经网络方法对水稻叶片色素含量进行高光谱反演。结果表明,BD与BP结合的估算模型对水稻叶片中的类胡萝卜素含量估算精度最高(R2=0.61,RMSEP=0.128 mg·g-1),BNA与BP结合的估算模型对水稻叶片中的叶绿素含量估算精度最高(R2=0.73,RMSEP=0.343 mg·g-1)。对比分析BDA与BP结合的模型和植被指数最佳回归模型的精度,发现波段深度分析建立的BP神经网络模型能较好地解决饱和问题,提高水稻叶片色素含量的估算精度。3.基于大田地面调查点和小区试验田的水稻灌浆期冠层反射光谱,根据HJ-1A卫星影像的光谱响应函数,模拟HJ-1A卫星的蓝、绿、红、近红外波段。对12个植被指数与水稻LAI的相关性进行分析,选取相关系数最大的植被指数构建水稻LAI估算模型。结果表明,GRVI的二次多项式回归模型估算水稻灌浆期LAI的精度最高,模型为LAI=-0.027*GRVI2+1.125*GRVI+0.028,建模集和验证集的R2分别达到0.89和0.80,RMSEC和RMSEP均较低,RPD大于2,模型优异。利用GRVI-LAI估算模型,获取水稻灌浆期LAI的空间分布图。由于HJ-1A卫星影像受大气状况和空间分辨率较低所限,空间分布图上的LAI预测值普遍低于地面对应点的LAI实测值。
[Abstract]:Rice is one of the main grain crops. Monitoring of its growth condition plays an important role in rice growth management, yield prediction, and disaster prevention and control. The leaf area index, pigment content and other physiological and biochemical parameters are important indicators of rice growth monitoring. The advantages of rich and so on can provide real-time, rapid, efficient and nondestructive rice growth, nutritional status and its change status, and provide technical support and theoretical basis for the information management of precision agriculture. In this paper, a canopy reflectance spectral inversion model of rice physiological and biochemical parameters was built with different nitrogen fertilizer levels and different growth stages. Based on the environmental satellite (HJ-1A) image data, the spatial distribution of the leaf area index (Leaf area index, LAI) of the critical growth period of rice was obtained, and the growth of rice was observed in a wide range. The following main results were carried out around the above content. 1. the main results were as follows: 15 vegetation indices were obtained based on the canopy reflectance spectrum data of rice. 5 traditional regression model analysis (linear function, exponential function, logarithmic function, power function, polynomial function) were carried out by the method of cross validation. The LAI hyperspectral index estimation model of rice at different growth stages was established, the optimum vegetation index was obtained at different growth stages, and the Noise Equivalent (NE) was used for the inversion of LAI into the vegetation index. The results showed that the normalized vegetation index (normalized difference vegetation index, NDVI) at the tillering stage, and the new vegetation index (new vegetation index, NVI) were sensitive to the change of LAI, and the accuracy was high, and the green normalized vegetation index (green normalized) at jointing stage, and the ratio vegetation index. Ratio vegetation index (RVI-3), the improved simple ratio index (modified simple ratio index, MSR) has Gao Min sensibility and estimation accuracy; the later GNDVI of the rice growth, the revised normalized difference vegetation index (modified normalized) is more applicable than the other indices. The traditional regression model of rice LAI in the whole growth period is too low, it is difficult to estimate the whole growth period rice LAI with unified vegetation index. Using partial least square regression modeling method, the decision coefficient R2 of modeling set and verification set can reach 0.87 and 0.81, RMSEC is 0.612, RMSEP is 0.856, RPD is greater than 2, and the whole life can be estimated more accurately. The rice LAI.2. is based on the hyperspectral data (400~750 nm) of the rice canopy based on the continuous removal of the rice (band depth, BD), the band depth ratio (band depth ratio, BDR), the normalized band depth (normalized band) and the normalized area band index (nm). Index. On this basis, principal component analysis (PCA) is used to achieve spectral dimensionality reduction, and then inverse propagation (back propagation, BP) neural network method is used for hyperspectral inversion of rice leaf pigment content. The results show that the estimation model of the combination of BD and BP is used to estimate the carotenoid content in rice leaves The accuracy is highest (R2=0.61, RMSEP=0.128 mg. G-1). The estimation accuracy of the estimation model combined with BNA and BP is the highest (R2=0.73, RMSEP=0.343 mg g-1). The accuracy of the model and the optimal regression model of the vegetation index is compared and analyzed. In order to solve the problem of saturation, the estimation accuracy of rice leaf pigment content was improved.3. based on the canopy reflectance spectra of rice grain filling stage in field ground survey point and plot test field. According to the spectral response function of HJ-1A satellite images, the blue, green, red and near infrared bands of the HJ-1A satellite were simulated. The correlation between the 12 vegetation indices and rice LAI was analyzed. The estimation model of rice LAI was constructed with the largest correlation coefficient of vegetation index. The results showed that the two polynomial regression model of GRVI was the highest for estimating the precision of LAI in rice filling period. The model was LAI=-0.027*GRVI2+1.125*GRVI+0.028, the R2 of modeling set and verification set reached 0.89 and 0.80 respectively, RMSEC and RMSEP were lower, RPD was greater than 2, and the model was excellent. The GRVI-LAI model was used to obtain the spatial distribution map of LAI during the grain filling period. Because the HJ-1A satellite image was limited by the atmospheric condition and the low spatial resolution, the prediction value of the LAI on the spatial distribution map was generally lower than the measured value of the LAI at the ground corresponding point.
【学位授予单位】：华中农业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S511;TP751

【参考文献】