高光谱与多光谱遥感水稻估产研究

发布时间：2018-01-22 08:56

本文关键词： 植被指数纹理特征随机森林水稻估产　出处：《武汉大学》2017年硕士论文　论文类型：学位论文

【摘要】：水稻是中国最主要的粮食作物,及时了解、准确掌握水稻种植面积、产量等对各级政府制定农业生产和农村政策、确保我国乃至世界粮食安全具有重要的意义。随着高光谱遥感技术和无人机遥感的发展与应用,对水稻估产的速度与精度都提出了更高的要求。本文利用水稻冠层尺度上的高光谱数据和低空无人机平台上获取的影像数据,对水稻产量反演的方法展开了相应的研究。本文通过观测武穴梅川镇实验区内的水稻,利用地面高光谱数据和无人机平台多光谱数据,提取植被指数、端元丰度和纹理特征等信息,使用逐步线性回归、BP神经网络和随机森林算法进行水稻估产研究,并对不同方法得到的不同模型精度进行了分析。主要研究内容如下:(1)针对高光谱数据和无人机多光谱影像各自的特点,利用敏感性分析和相关性分析,分别提取了2种地面高光谱植被指数、4种无人机多光谱植被指数、5种端元丰度和4种纹理特征等因子。结果表明:地面高光谱植被指数中与水稻产量相关性最高的指数为孕穗期DVI(431,665)和乳熟期DVI(892,1087)相关系数值分别为0.77和0.76;从不同生育期无人机多光谱影像获取的植被指数、端元丰度和纹理特征3类参数中与水稻产量相关性最高的是抽穗期NDVI指数和孕穗期CIrededge指数,相关系数分别为0.795和0.781;端元丰度中与水稻产量相关性最高的是抽穗期叶片丰度和稻穗丰度,分别为0.695和-0.696;纹理特征与水稻产量相关性最低。(2)利用筛选出的特征因子,采用"留一法"交叉验证的方式分别使用逐步线性回归、BP神经网络和随机森林算法进行建模。结果表明:当利用地面高光谱孕穗期DVI和乳熟期DVI作为自变量时,得到的水稻估产模型R2依次为0.633、0.575 和 0.440,对应的 RMSE 分别为 24.546 kg/亩、27.710 kg/亩和 29.704 kg/亩;当输入变量为无人机植被指数、端元丰度和纹理特征时,不论是用逐步线性回归、还是BP神经网络或者随机森林算法,植被指数的表现都比端元丰度和纹理特征要好,再次表现出了植被指数的强大,3种研究方法中建模效果最好的均为以植被指数作为自变量的模型,R2分别是0.615、0.599和0.405,RMSE分别为 25.131 kg/亩、26.377 kg/亩和 30.628 kg/亩。(3)综合地面和低空无人机数据中获取的多种特征,利用主成分分析法进行主成分提取,通过逐步线性回归得到的多平台综合估产模型相对于地面或无人机平台的最优模型来说,效果有一定提高,利用交叉验证的方法进行建模时,平均R2达到 0.661,平均 RMSE 为 23.568 kg/亩。
[Abstract]:Rice is the most important food crop in China. Timely understanding and accurate understanding of rice planting area, yield and other government at all levels to formulate agricultural production and rural policies. With the development and application of hyperspectral remote sensing technology and UAV remote sensing technology, it is of great significance to ensure food security in China and the world. Higher requirements for the speed and accuracy of rice yield estimation are put forward. In this paper, the hyperspectral data on the rice canopy scale and the image data obtained on the low altitude UAV platform are used. Based on the observation of rice in the experimental area of Meichuan Town in Wuxuan, the vegetation index is extracted by using the ground hyperspectral data and the multi-spectral data of UAV platform. The yield estimation of rice was studied by stepwise linear regression BP neural network and stochastic forest algorithm. The main research contents are as follows: (1) according to the characteristics of hyperspectral data and UAV multi-spectral images, sensitivity analysis and correlation analysis are used. Two kinds of ground hyperspectral vegetation index and four kinds of UAV multi-spectral vegetation index were extracted. The results showed that the highest correlation between surface hyperspectral vegetation index and rice yield was DVI(431 at booting stage. The correlation coefficient of 665) and milk stage DVI892 (1087) were 0.77 and 0.76, respectively. Vegetation index obtained from multi-spectral images of UAV at different growth stages. The NDVI index at heading stage and CIrededge index at booting stage were the highest among the three kinds of parameters, the correlation coefficients were 0.795 and 0.781respectively. Leaf abundance and panicle abundance at heading stage were the highest correlation between end element abundance and rice yield, which were 0.695 and -0.696, respectively. Texture features have the lowest correlation with rice yield. BP neural network and stochastic forest algorithm were used to model the model. The results showed that when the ground hyperspectral booting stage DVI and milk-ripening stage DVI were used as independent variables. The rice yield estimation model R2 was 0.633C 0.575 and 0.440 respectively, and the corresponding RMSE was 24.546 kg / mu. 27.710 kg / mu and 29.704 kg / mu; When the input variables are UAV vegetation index, Duan Yuan Feng degree and texture feature, whether by stepwise linear regression, BP neural network or stochastic forest algorithm. The performance of vegetation index is better than Duanyuanfeng degree and texture feature. Again, the model with vegetation index as independent variable is the best modeling method among the three research methods of vegetation index. The RMSE of RMSE was 25.131 kg / mu, respectively. 26.377 kg / mu and 30.628 kg / mu. Compared with the optimal model of the ground or UAV platform, the multi-platform comprehensive yield estimation model obtained by stepwise linear regression is more effective than the optimal model of the ground or UAV platform, when the cross-validation method is used to model the model. The average R2 was 0.661 and the average RMSE was 23.568 kg / mu.
【学位授予单位】：武汉大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S511;TP79

【相似文献】