利用CARS-BPNN模型的南疆枣园土壤有机质高光谱反演

Hyperspectral Inversion of Soil Organic Matter in Jujube Orchardin Southern Xinjiang Using CARS-BPNN

土壤有机质(SOM)含量是制定枣园土壤施肥方案的主要依据。合理的施肥方案对提升红枣品质、减少农户投入和增加枣园产出有重要意义。利用传统方法获取枣园SOM含量耗费时间和资源,不符合枣园精准施肥管理的需求,土壤有机质高光谱检测是一种有效的替代方法。为筛选南疆枣园SOM的高光谱快速检测模型,采用网格布点法采集158个枣园土壤样品,测定风干土样的室内高光谱数据和SOM含量。分别将400~2400 nm全波段(R)和通过竞争自适应加权算法(CARS)、连续投影算法(SPA)、粒子群优化算法(PSO)三种数据降维算法筛选的数据集与偏最小二乘回归(PLSR)、反向传播神经网络(BPNN)、卷积神经网络(CNN)三种建模方法结合构建12种枣园SOM含量的组合反演模型,通过对比模型的精度评价指标和训练时间,筛选枣园SOM含量最优光谱反演模型。结果表明:(1)CARS、SPA、PSO三种降维算法都能将光谱数据压缩至原来的10%以下,筛选波长数分别由原来的2001个变量降为98、156、102个,降维组合模型的验证集RPD均大于1.50,均能实现对枣园SOM含量的反演,与R组合模型相比,降维组合模型至少能节省30%的时间成本,特别是与BPNN和CNN等构建的组合模型,能节省90%的训练时间,且模型稳定性更强,模型效果更优。(2)CARS数据集构建组合模型的验证集R^(2)均大于0.85,RPD均大于2.50,RPIQ均大于1.60,在三种降维算法中效果最好;PSO数据集的组合模型验证效果略低于CARS数据集,但优于R数据集,R^(2)均大于0.80、RPD均大于2.00;SPA数据集构建组合模型的验证效果要低于R数据集,在三种降维算法中效果最差。(3)BPNN和CNN两种方法的反演模型验证效果均优于PLSR模型,而在模型训练时间和模型验证效果等方面,BPNN模型优于CNN模型,其结合CARS数据集的验证效果最优,R^(2)为0.91、PRD为3.34、RPIQ为3.17、nRMSE%为11.93,训练时间为58.00 s,模型符合快速检测枣园SOM含量的要求。CARS-BPNN模型为反演南疆枣园SOM的最优模型,研究结果能够为南疆枣园土壤养分快速检测与制定施肥方案提供参考。

The soil organic content is the main basis for developing soil fertilization programs in jujube orchards.A reasonable fertilization program is of great significance for improving the quality of jujube,reducing farmers investment and increasing the output of jujube orchards.However,it is time-consuming and resource-intensive to obtain SOM content of jujube orchards using the traditional method,which does not meet the needs of precise fertilization management in jujube orchards.At the same time,the hyperspectral detection of soil organic matter is an effective alternative method.158 soil samples are collected by grid distribution method,and the indoor hyperspectral data and SOM content of air-dried soil samples are determined.The 400~2400 nm full waveband(R)and the datasets selected by three data reduction algorithms of competitive adaptive weighting algorithm(CARS),successive projection algorithm(SPA)and particle swarm optimization algorithm(PSO)are combined with three modeling methods,which are partial least squares regression(PLSR),back propagation neural network(BPNN)and convolutional neural network(CNN)to construct 12 combined inversion models of SOM content of jujube orchards.Moreover,the optimal spectral inversion model of SOM content of jujube orchards was selected by comparing the accuracy evaluation index and training time of the models.The results show that(1)CARS,SPA and PSO can all compress the spectral data to less than 10%of the original data,and the number of screened wavelengths is reduced from the original 2001 variables to 98,156 and 102,respectively.The validation set RPD of the dimensionality reduction combined model are all greater than 1.50,and all of them can achieve the inversion of the SOM content of jujube orchards.Compared with the R combined model,the dimensionality reduction combined model can save at least 30%of time cost,especially the combined model constructed with BPNN and CNN can save 90%of the training time,and the model has stronger stability and better model effect.(2)The validation set of the CARS dataset to construct the combined model has R^(2)greater than 0.85 and RPD greater than 2.50,which is the best among the three-dimensionality reduction algorithms;the validation effect of the combined model of the PSO dataset is slightly lower than that of the CARS dataset,but better than that of the R dataset,with R^(2)greater than 0.80 and RPD greater than 2.00;the validation effect of the SPA dataset to construct the combined model is lower than that of the R dataset The validation effect of the SPA dataset is lower than that of the R dataset,and the effect is the worst among the three-dimensionality reduction algorithms.(3)Both BPNN and CNN methods outperformed the PLSR model in terms of inversion model validation,while the BPNN model out performed the CNN model in terms of model training time and model validation effect,and its validation effect combined with the CARS dataset is optimal with R^(2)of 0.91,PRD of 3.34,nRMSE%of 11.93,and training time of 58.00 s.The model can detect the SOM content of the jujube orchards rapidly.The CARS-BPNN model is the optimal model for the inversion of SOM in jujube orchards in South Xinjiang,and the results of the study can provide a reference for the rapid detection of soil nutrients and formulation of fertilization plan in jujube orchards in South Xinjiang.