黑土有机质含量的高光谱估测模型研究

发布时间：2018-02-27 16:17

本文关键词： 黑土有机质高光谱多元逐步回归偏最小二乘回归支持向量机　出处：《吉林大学》2015年硕士论文　论文类型：学位论文

【摘要】：东北黑土是不可再生的宝贵资源，关系到东北地区乃至全国的粮食生产、经济发展以及生态环境，然而对黑土的过度开发以及不合理的耕作方式等造成了土壤的贫瘠化，水土流失严重。因此有必要及时对土壤属性进行测定与质量评价，保护黑土资源刻不容缓。高光谱遥感技术的发展为土壤理化参数的获取提供了一种便捷、高效的手段。相比传统方法，更省时省力，降低成本；相比常规多光谱遥感，能提高反演精度。其较高的光谱分辨率能为分析土壤表层情况、内在属性等提供详细的数据，目前已经在估测土壤有机碳、水分、氮、磷、钾等相关元素含量方面有较为广泛的应用。本文以东北典型黑土区的农田土壤为研究对象，以68个黑土样本的实测高光谱数据和有机质含量为主要数据源。土壤光谱曲线是其各个属性成分综合作用的结果，分析土壤的反射光谱特征是进行有机质含量估测的基础。在对黑土光谱数据进行断点校正、平滑去噪和变换处理后，首先比较分析了不同有机质含量等级下、不同光谱分辨率下，黑土室内光谱曲线特征的差异，并将室内、野外黑土光谱进行了对比，利用单相关分析，得出有机质的响应波段范围。然后，基于光谱分析技术和统计学原理，利用黑土光谱数据及其变换形式分别建立了多元逐步回归、偏最小二乘和支持向量机回归模型来对有机质含量进行估测，并分析对比了各种模型的精度。主要得到以下结论：黑土室内光谱曲线整体光滑平缓，反射率随着波长增大而逐渐升高，属于有机质控制类型，有机质影响着整个波段的光谱特征，尤其是可见光波段。室内与野外的黑土光谱曲线特征基本相似，由于不受外在环境影响，，光谱曲线更加光滑规整，一些细节特征更加明显。有机质响应波段范围较宽，当采样间隔增至10nm时，曲线更为光滑，光谱特征几乎没有改变；当采样间隔大于20nm时，黑土光谱的吸收特征随着光谱分辨率的降低逐渐削弱。有机质含量与黑土光谱反射率呈负相关，显著相关波段为550～680nm，吸光度一阶微分变换对增强相关性的作用最大，在1274nm处，相关系数达到-0.8。在多元逐步回归模型中，吸光度一阶微分及连续统去除一阶微分的建模效果较好，将黑土有机质含量的对数形式作为因变量后，每个模型的建模和预测精度都有提高。利用全光谱数据建立的偏最小二乘回归模型，建模结果较多元逐步回归有很大改善，其中吸光度光谱一阶微分的PLSR模型结果最优，RPD为2.664，有较好的预测能力。不同光谱分辨率下的PLSR模型，随着光谱分辨率的降低，精度先提高后逐渐降低，在10nm采样间隔时，模型效果最好，说明重采样后不仅保留了原始的光谱特征，还减少了数据的冗余，削弱了噪声的影响。最小二乘支持向量机回归建模结果要优于前两种线性模型，特别是通过PLS提取特征主成分后，不仅缩短了计算时间，模型稳定性和预测能力都得到增强，R2C为0.993，RPD为3.682，缩小了建模和预测精度的差距。
[Abstract]:Northeast China is a valuable non renewable resources, related to the northeast region and the country's grain production, economic development and ecological environment, however, over exploitation of black soil and unreasonable tillage caused the impoverishment of soil, the soil erosion is serious. Therefore it is necessary to timely determination and quality evaluation of soil properties, protection the black soil resources is urgent. For the development of hyperspectral remote sensing technology for soil physicochemical parameters provides a convenient and efficient method. Compared with the traditional method, more time-saving, reduces the cost; compared with the conventional multi spectral remote sensing, can improve the inversion precision. Its high spectral resolution for analysis of surface soil, such as intrinsic property provide detailed data, has been in the estimation of soil organic carbon, water, nitrogen, phosphorus, is widely used on related elements content of potassium.
In this paper, farmland soil in typical black soil region of Northeast China as the research object, based on the measured hyperspectral data of 68 soil samples and organic matter content as the main data source. Soil spectral curve is the comprehensive result of each attribute composition, spectral characteristics analysis is based on estimation of soil organic matter content in the soil. The spectral data breakpoint correction, denoising and transform processing, the first comparative analysis of different organic content levels, different spectral resolution, different spectral characteristics of black soil indoors, and the indoor, compared with wild black soil spectra, using single correlation analysis, the organic matter and response wave band range. Spectral analysis and statistics, based on the principle of multiple stepwise regression were established using soil spectral data and the transformation form, the partial least squares regression model and support vector machine The content of organic matter is estimated, and the accuracy of various models is analyzed and compared. The following conclusions are obtained:
The indoor soil spectral curve overall smooth flat reflectivity with the wavelength gradually increases, which belongs to the type of organic matter control, organic matter affects the spectral characteristics of the band, especially the visible characteristics of black soil. The spectral curve of indoor and field similar to that by not subject to the impact of the external environment, spectral curve is more smooth, some the details of the more obvious characteristics of the organic matter. In response to a wide range of bands, as the sampling interval is increased to 10nm, the curve is more smooth, almost no change in spectral characteristics; when the sampling interval is greater than 20nm, soil absorption spectrum characteristics with reduced spectral resolution is gradually weakened. Organic matter content and soil reflectance was negatively correlated significantly the relevant band is 550 ~ 680nm, the absorbance of first order differential transform to enhance the relevance, at 1274nm, the correlation coefficient reached -0.8.
In the multivariate regression model, the first-order differential absorbance and continuum removal effect of modeling first-order differential is better, the logarithm of soil organic matter content as the dependent variables, modeling and prediction accuracy of each model are improved. Partial least squares regression model is established by using optical spectrum data, the modeling results are multiple stepwise regression is greatly improved, which results in the absorbance spectra of first order differential optimal PLSR model, RPD 2.664, has a good predictive ability. The PLSR model of different spectral resolution, with reduced spectral resolution, precision first increased and then decreased gradually, 10nm in the sampling interval, the best model, that not only re sampling retain the spectral features of the original, but also reduces the data redundancy, weaken the effect of noise. The least squares support vector machine regression modeling results are better than those of the first two kinds of linear models, especially After extracting feature principal components from PLS, the computation time is shortened, and the stability and prediction ability of the model are enhanced. R2C is 0.993 and RPD is 3.682, which reduces the gap between modeling and prediction accuracy.

【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：S153.6

【参考文献】