Research On Grassland Forage Hyperspectral Image Recognition Algorithms Based On Machine Learning

In recent years,the grassland forage survey and monitoring work is mainly based on satellite remote sensing multispectral images,but it is restricted by factors such as overall resolution,sampling period,cost and so on.However,the hyperspectral image of high-resolution acquired in close range can solve the above faults.And it is rarely studied at present.Therefore,this paper uses grassland forage hyperspectral images as the study object.Based on existing machine learning algorithms,we proposed classification algorithms to improve the accuracy and efficiency of pasture recognition and lay ground for grassland intelligent management.The main research contents and conclusions are as follows:(1)Collecting and establishing the data set of grassland forage hyperspectral images.The grassland forage hyperspectral image with visible-near infrared spectrum(400nm～1000nm)was collected by using a hyperspectral instrument(Hyper Spec(?)PTU-D48E)in the field.There are 16 species of forage.Considering the influence of the difference in the spectral characteristics of different growth periods and growth regions of forage,we established the HSI data set of the same growth period in the same growth area and the HSI data set of different growth periods in different growth areas,respectively.(2)Data preprocessing.In the process of collecting images in the field,there is noise and distortion,due to the growth status of the forage samples is unevenly distributed,the plant heights are different,and susceptible to interference from external natural factors,such as light,wind,climate,shadows,etc..So we need to preprocesse the data.Firstly,we extracted the ROI from collected images.Then,using MSC,SNV,Normalization,S-G,and Nirmaf for preprocessing.Among them,MSC is the best that effectively eliminates the noise,and guarantees the accuracy and stability of the prediction model.(3)We proposed classification models suitable for grassland forage hyperspectral image.Model 1: We constructed a forage hyperspectral data classification model based on RBF-SVM.Firstly,we used the PCA white for feature extraction;then,we selected the appropriate principal component and input to SVM;finally,we optimized the experiment through K-fold cross-validation.Comparing the automatic classification results of grassland forages hyperspectral images corresponding to different SVM kernel functions.Among them,the RBF-SVM classification result was better,which the OA is over 95%.The results show that MSC-PCA-SVM(RBF)has simple structure,easy to train,fast in convergence,and has high classification accuracy.RBF-SVM compared with GBDT and KNN,that it has certain advantages and is an efficient and robust recognition model.Model 2: We constructed a forage hyperspectral image classification model based on variance selection and gaussian naive bayes.Firstly,we proposed a feature extraction method of PCA white based on variance selection(V-pcaw)method.We optimized the effective information in features by the dimensionality reduction method based on variance selection,and selected the best feature according to the threshold and principal components.The number of variables was 2;then,we used Gaussian NB and SVM in combination with K-fold cross-validation to establish recognition models,respectively;finally,we evaluated the model by Kappa coefficient,OA,test time and other indicators.Comparing V-pcaw with PCA,which V-pcaw’s average of OA and Kappa coefficient were increased by 2.995% and 0.05025,respectively.Comparing the Gaussian NB model and the SVM model under the same circumstances,the Gaussian NB model consumes less time while ensuring better classification accuracy.Among them,the MSC-V-pcaw-Gaussian NB model has the best recognition effect and takes the least time.The values of OA,Kappa coefficient and test time are 99.33%,0.99 and 0.002022 s,respectively.The results show that the method based on variance selection and gaussian naive bayes can get better feature expression of pasture hyperspectral images and achieve better recognition performance.Model 3: We constructed a recognition model of forage hyperspectral image based on F-SVD and XGBoost.Aiming at the issues of high time complexity and poor accuracy in the application of SVD in hyperspectral recognition,we proposed an improved SVD algorithm,F-SVD,which introduced latent factors(F)into the SVD decomposition strategy for improved singular matrices,thinking of the correlation between latent variables and original variables.Then,we established a forage recognition model based on XGBoost.F-SVD-XGBoost compared with FA-XGBoost and SVD-XGBoost,F-SVD-XGBoost has increased OA by 1.98% and 1.67%,and reduced time consumption by 1.369 s and 0.522 s,respectively.The results show that F-SVD-XGBoost has obvious advantages in terms of time wastage and classification accuracy,especially for multi-category forage samples.