基于遗传算法结合偏最小二乘的潮土碱解氮高光谱特征及含量估测

提取土壤碱解氮特征光谱是利用高光谱数据进行其含量估测的关键。对山东省典型潮土土壤样本测试高光谱并进行变换；采用遗传算法（GA）结合偏最小二乘法（PLS），在筛选潮土碱解氮含量特征谱区的基础上，构建潮土碱解氮含量偏最小二乘（PLS）回归估测模型；优选最佳模型并与相关分析、逐步回归分析和单纯偏最小二乘回归分析的模型进行比较。结果表明：潮土碱解氮特征波段为449~469nm，988~1001nm，1065~1078nm，1716~1736nm，1912~1925nm，2213~2233nm，2262~2275nm；基于各输入光谱特征谱区构建的估测模型决定系数R²均较高，其中基于反射率一阶导数光谱筛选的特征谱区，构建的模型精度最高，数据点（147个）为原始全谱的7.17%，建模R²达到0.97，均方根误差RMSE为4.78mg/kg，验证R²为0.95，RMSE为5.49mg/kg，对潮土碱解氮含量具有较好的预测准确性；在光谱变换形式中，反射率的一阶导数表现最佳；方法比较显示采用遗传算法结合偏最小二乘（GA-PLS）获得较高预测精度的同时，可简化模型。说明遗传算法结合偏最小二乘法（GA-PLS），可有效筛选土壤碱解氮的特征波段，减少模型参与变量，提高估测精度。

Hyperspectral Characteristic and Estimation Modeling of Fluvo-aquic Soil Alkali Hydrolysable Nitrogen Content Based on Genetic Algorithm in Combination with Partial Least Squares

Extracting the characteristic spectra is the key of estimating soil alkali hydrolysable nitrogen content based on hyperspectra. This article was carried out to detect and transform the hyperspectra reflectance of the soil samples collected from the fluvo-aquic soil areas in Shandong Province. The characteristic wave bands of fluvo-aquic soil alkali hydrolysable nitrogen were filtered using genetic algorithms and partial least squares method. Then, the partial least squares regression estimation models of fluvo-aquic soil alkali hydrolysable nitrogen content were constructed. The best model was selected and compared with the models based on correlation analysis, stepwise linear and partial least squares regression directly. The results showed that the characteristic wave bands of fluvo-aquic soil alkali hydrolysable nitrogen were 449-469 nm, 988-1001 nm, 1065-1078 nm,1716-1736 nm, 1912-1925 nm, 2213-2233 nm and 2262-2275 nm, the modeling coefficient of determination(R2) based on characteristic wave bands of different input spectra was generally high, the model based on the characteristic wave bands of the first derivative of reflectance had the highest precise, with the 147 data points as 7.17% of the original spectrum, the model building R2 to 0.97, RMSE as 4.78 mg/kg, the validated R2 to 0.95, RMSE as 5.49 mg/kg, the model had good prediction accuracy of fluvo-aquic soil alkali hydrolysable nitrogen content. The comparison of methods showed that the genetic algorithms combined with partial least squares regression could not only obtain higher prediction accuracy, but also simplify models. Therefore, the genetic algorithms combined with partial least squares regression method may effectively select the characteristic wave bands of soil alkali hydrolysable nitrogen, reduce the model variables and improve the estimation accuracy.