Hyperspectral Detection Of Black Soil Organic Matter Classification Based On SA-Double-Stacking Algorithm

In order to improve the accuracy of the black soil organic matter classification detection model,realize the accurate management of zoning,promote the implementation of variable fertilization,meet the needs of precision agriculture,and at the same time solve the problems of the existing model is not targeted,organic matter discrimination model is not targeted,data pre-processing is complicated,and the ability to build model generalization is weak,etc.This paper collects the physical and chemical properties and the hyperspectral data set of the soil in the Xiangyang experimental base and proposes a model for black soil organic matter content In this paper,we collected the physical and chemical properties and hyperspectral data sets from Xiangyang experimental site,and proposed an SA-LSVM-DS model for organic matter content of black soil.(1)Collection of black soil samples and pre-processing of spectral data.After completing the field sampling,chemical analysis,and spectral data collection of black soil samples;completing the deletion of missing spectral data by data cleaning check;using SG smoothing to pre-process the data,dividing the samples into training set,validation set,and test set to obtain the sampled complete true value-spectral data set.(2)Simulated Annealing(SA)is introduced as the hyperparameter optimization algorithm for the hyperparameter preference scheme.Compared with the widely used grid search and random search,the simulated annealing algorithm has a stronger data concentration and interval aggregation effect and has a higher hyperparameter optimization efficiency.(3)A structural improvement of the machine learning model is designed based on the Stacking algorithm,and a Double-Stacking structural model is obtained.Through the integration and encapsulation of the primary learner in the L1 level,the learning method is introduced instead of the traditional equal-weighted voting scheme,and a learnable iterative training model is realized with the probabilistic output of the L1 level,which reduces the degree of influence produced by the incorrect prediction labels of the primary learner on the L2 level learner and effectively improves the prediction accuracy of the model.(4)Experimentally,nine Double-Stacking models were constructed and hyperparametric optimization analysis was performed on the nine models.In the comparison,the SA-DS model with excellent generalization degree,the SA-LSVM-DS model,was selected.Further,the SA-LSVM-DS model proposed in this study is compared with ten SOTA models commonly used for hyperspectral classification,and combined with the generalized hyperparametric optimization scheme,based on comparing the accuracy before and after optimization and the accuracy on the test set,it can be proved that the SA-LSVM-DS model proposed in this paper is the best in black soil organic matter classification detection.(5)The feature-band analysis was completed for the black soil hyperspectral data using the XGBoost algorithm,and the feature data modeling work was completed.The 34 characteristic bands with the top 30% weight were selected for band reorganization of the dataset,the LSVM-DS model and ten SOTA comparison models were re-fitted and hyperparameter optimization processes were completed using the characteristic band dataset for comparative analysis.By comparing the evaluation indexes of each model,the SALSVM-DS model proposed in this study,compared with other single-model schemes,can still have good and stable application performance without complex feature band selection.Through the above study,the SA-LSVM-DS model was developed for the classification detection of black soil organic matter based on hyperspectral technology,which has an accuracy of 87.60% and 94.88% in the validation and test sets,which is 38.65% higher than the accuracy of the unoptimized LSVM-DS model.Compared with other comparable models,the accuracy in the validation set improved by 6.31%～28.33%,the accuracy in the test set improved by 11.42%～23.19%,and the recognition accuracy of soil samples at all levels were improved.the SA-LSVM-DS model has the strongest generalization ability while meeting the accuracy requirements,and it can directly analysis of full-band data,realizing the process feasibility of single sample detection,solving the problem of low efficiency of traditional chemical detection,and helping to complete the study of precision management zoning,promoting the implementation of variable fertilization and achieving the goal of precision agriculture.