基于水稻高光谱遥感数据的植被指数产量模型研究

发布时间：2018-04-09 12:24

  本文选题：植被指数遥感估产　切入点：高光谱　出处：《沈阳农业大学》2017年硕士论文

【摘要】：我国主要的粮食产物为水稻,由于它有价格低廉,易于培育,产量稳定等优点,因此国家和政府对水稻的产量问题高度重视,水稻产量的准确预测也对我国粮食安全有重要意义。由于传统计算水稻面积的方法耗费人力,耗费物力,耗费财力,并且不能对水稻进行大面积、全天的实时监测,基于传统水稻方法的缺点,遥感估产技术应运而生,它弥补了传统估产的缺陷,并且对数据的监测更加迅速及时范围广。无人机的发展给遥感技术又提供了一个新的实现手段,近年来,利用无人机搭载高光谱或者多光谱遥感估产掀起了热潮。本文根据2015年的光谱数据,研究辽宁省沈阳农业大学一块试验田的水稻产量,利用不同的方法建立了植被指数与产量的模型,得到可靠的精确度和稳定性,对日后研究水稻产量估算的发展提供借鉴和指导。本文首先利用二元定距变量分析的方法对单个NDVI和各时期像元之和和产量进行了相关性分析,筛选较好的NDVI因子分别利用水稻冠层NDVI数据建立一元、多元和主成分分析后的回归模型,其结果相对比较理想。通过比对各个模型的参数,最终得出分蘖期和抽穗期的组合复合模型要比单天复合和各生长时期的复合模型拟合度高,其R2达到0.805,F值为9.280,显著性为0.008,其相对误差和均方根误差分别为5.06%和324kg/hm2,具有较高的精确度,且最佳估产时间为分蘖期和抽穗期;主成分分析所得的模型R2达到0.806,F值为20.789,显著性为0,结果更为精确。然后用同样的方法建立PRI与产量的回归模型,在各个复合模型和主成分分析回归模型中,利用水稻整个生长时期建立的复合估产模型是最理想的,其R2达到0.875,F值为14.965,显著性为0.011,经过检验其相对误差和均方根误差分别为3.89%和402kg/hm2,具有较高的精确度。最后将两种指数结合进行产量建模,分别使用了多元线性回归、主成分分析和神经网络算法进行估产。对于多元线性回归所得模型,其R2=897,F=10.701,显著性为0.021,此模型比单独使用NDVI或者PRI所得模型拟合度都高;主成分分析所得模型的R2=922,F=19.567,显著性为0.003;对于神经网络算法所得的模型,其拟合度达到0.959,其预测精度也比前面的模型高,相对误差和均方根误差分别为1.31%和66kg/hm2,所以BP神经网络是最理想的预测水稻产量的方法。
[Abstract]:The main grain product in China is rice. Because it has the advantages of low price, easy cultivation and stable yield, the government and the state attach great importance to the problem of rice yield.The accurate prediction of rice yield is also of great significance to food security in China.Because the traditional method of calculating rice area consumes manpower, material resources and financial resources, and cannot monitor rice in real time all day, based on the shortcomings of traditional rice method, remote sensing yield estimation technology emerges as the times require.It makes up for the shortcomings of traditional yield estimation, and the monitoring of data is more rapid and timely.The development of unmanned aerial vehicle (UAV) provides a new method for remote sensing technology. In recent years, the use of unmanned aerial vehicle (UAV) to carry hyperspectral or multispectral remote sensing has raised a boom.Based on the spectral data of 2015, the yield of rice in a field of Shenyang Agricultural University of Liaoning Province was studied. The models of vegetation index and yield were established by using different methods, and the reliable accuracy and stability were obtained.It provides reference and guidance for future research on rice yield estimation.In this paper, the correlation analysis of the sum and yield of single NDVI and each period pixel was carried out by using the method of binary fixed distance variable analysis, and the better NDVI factors were selected to establish a single element by using the rice canopy NDVI data, respectively.The regression model after multivariate and principal component analysis is relatively ideal.By comparing the parameters of each model, it was concluded that the combination model at tillering stage and heading stage had higher fitting degree than that at single day and each growth stage.The R2 value was 9.280, and the significance was 0.008. The relative error and root mean square error were 5.06% and 324kg / hm ~ 2, respectively, which had high accuracy and the best estimation time was tillering stage and heading stage.The model R2 obtained by principal component analysis (PCA) was 0.806F (20.789), and the result was more accurate.Then, the regression model of PRI and yield was established by the same method. In each compound model and principal component analysis regression model, the compound yield estimation model established by using the whole growth period of rice was the most ideal.The R2 value was 14.965 and the significance was 0.011.The relative error and root mean square error were 3.89% and 402kg / hm ~ 2, respectively, which showed high accuracy.Finally, the two indices are combined to model the yield, and the multiple linear regression, principal component analysis and neural network algorithm are used to estimate the yield.For the multivariate linear regression model, the R2O897FU 10.701 model has a significant difference of 0.021, which is higher than that obtained by using NDVI or PRI alone, and the R2O922FU 19.567model obtained by principal component analysis has a significance of 0.003. For the model obtained by neural network algorithm,The relative error and root mean square error are 1.31% and 66kg / hm ~ 2, respectively. So BP neural network is the best method to predict rice yield.
【学位授予单位】：沈阳农业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S511;S127
，

本文编号：1726435