基于核函数与可见光光谱的大豆植株群体净光合速率预测模型

Prediction Model on Net Photosynthetic Rate of Soybean Plant Groups Based on Kernel Function and Visible Light Spectrum

采用美国MSR-16便携式多光谱辐射仪,通过推导的辐射仪有效观测面积公式,计算出测试单元数量,有效解决了测量区域可见光各波段光谱辐射配比关系M_D所需测量次数不确定的难题。采用美国CID公司生产型号为CI-310便携式光合作用测定系统,测量大豆植株群体净光合速率C_D。通过[0,1]归一化方法对M_D和C_D进行归一化处理,分别得到归一化数据M_D_1和C_D_1。按不同测试时间划分,将M_D_1分成两部分数据M_D_(11)和M_D_(12),将C_D_1分成两部分数据C_D_(11)和C_D_(12)。使用polynomial核函数、gauss核函数、sigmoid核函数和自主研发的bio-selfadaption核函数,利用grid-search,Genetic Algorithm,Particle Swarm Optimization对支持向量机惩罚参数c和参数g寻优,在支持向量机epsilon-SVR公式、nu-SVR公式条件下,通过四种核函数、三种优化方法、两种公式的交叉组合、M_D_(11)、C_D_(11),建立大豆植株群体净光合速率预测模型。试验结果表明,在大豆植株试验区域面积S=17m^2和MSR-16便携式多光谱辐射仪放置于大豆植株冠层上方高度H=2m条件下,epsilon-SVR-bio-selfadaption-grid-search模型对预测集1 C_D_(12)的预测精度达到85%以上,对预测集2 C_D_(12)的预测精度达到82%以上。在S和H其他组合条件下,epsilonSVR-bio-selfadaption-grid-search模型对预测集2 C_D_(12)的预测精度达到81%以上。epsilon-SVR-bio-selfadaption-grid-search模型表明了bio-selfadaption核函数有效性、测量区域可见光光谱数据方法合理性、利用可见光光谱预测大豆植株群体净光合速率可行性。

The paper uses MSR-16 portable multispectral radiometer made in the USA and computes the numbers of the test units by pulling the formula on the radiometer effective observation area, which solves the problem on the uncertain numbers of computing the times on region visible light band spectral radiation ratio M_D. The paper uses CI-310 portable photosynthesis measurement system made by American CID Company and measures the net photosynthetic rate of a group of soybean plant. M D and C D are normalized by the normalization method [0, 1]. Then, the normalization data M_D1 and C_D1 are gained . Based on the different test time, M_D1 is divided of M Dll and M_D12. C_D1 is divided of C_DH and C_Dlz. The paper uses polynomial kernel function, gauss kernel function, sigmoid kernel function and bio-selfadaption kernel function constructed by us with Support Vector Machine. Penalty parameter c and parameter g separately are optimized with optimization algorithms such as grid-search,genetic algorithm and particle swarm optimization. Based on the formula epsilon-SVR and the formula nu-SVR with Support Vector Machine, the paper constructs the prediction model on the net photosynthetic rate of a group of soybean plant by using of the cross combination with four kernel functions, three optimization methods and two formulas. The test results are as follows： in the condition of S=17 m^2 which is the test plan area of soybean plant and the H=2 m which is the high on MSR-16 portable multispectral radiometer above the canopy of soybean plant, the prediction accuracy is up to 85% on the No. 1 prediction set C_D12 and the prediction accuracy is up to 82% on the No. 2 prediction set C_D12 based on the model epsilon- SVR-bio-selfadaption-grid-search. In the condition of other combinations with S and H, the prediction accuracy is up to 81% on the No. 2 prediction set C_D12 based on the model epsilon-SVR-bio-selfadaption-grid-search. The model epsilon-SVR-bio-selfada- ption-grid-search indicates the validity of bio-selfadaption kernel functions which is constructed by our previous research with support vector machine. The model epsilon-SVR-bio-selfadaption-grid-search indicates the rationality of the measure method on visible spectral data in the test area. The model epsilon-SVR-bio-selfadaption-grid-search indicates the feasibility of the prediction method on net photosynthetic rate of soybean plant groups by using of visible spectrum.