Research On Feature Extraction Method Based On Stacked Autoencoder And Its Applications

In recent years,the progress and rapid development of hyperspectral imaging technology has not only aroused great interest in the remote sensing field,but also achieved good achievements and breakthrough in the application research of precision agriculture,medical detection,resource exploration,environmental monitoring,disaster early warning,target reconnaissance and other civil and military fields.At present,the application of hyperspectral data presents a trend of processing,analysis and application of hyperspectral data based on artificial intelligence technology.The further breakthrough of hyperspectral imaging technology and the traction of application research demand will better promote the innovation and development in the application field of hyperspectral data.Based on hyperspectral data as the main research object,SAE-BP deep neural network model is proposed in this paper based on stacked autoencoder and BP neural network to solve feature extraction problems.In order to solve the two basic problems of regression prediction and classification in the field of academic research,SAE-BP model is established and successfully applied in the fields of soil temperature prediction and crop species recognition.The main content of this paper has three parts.On the basis of theoretical research on machine learning methods,combined with the characteristics and challenges of hyperspectral data,this paper conducts in-depth analysis and research on how to establish a more effective feature extraction model.The main work contents are as follows:(1)Firstly,the subject background of the hyperspectral data application and research purpose and significance are introduced,the research status and applications of machine learning algorithms in feature extraction at home and abroad are also analyzed,which both provide a theoretical basis for feature extraction methods based on stacked autoencoder.In this paper,the SAE-BP hybrid deep neural network model is proposed according to the characteristics and challenges of hyperspectral data,and the principle explanation and algorithm implementation are also carried out.SAE-BP model is a hyperspectral data feature extraction method based on stacked autoencoder and BP algorithm.It makes full use of the powerful feature extraction and describing ability of stacked autoencoder and the excellent self-learning and generalization ability of BP neural network,and combines unsupervised learning with supervised learning.In this paper,the SAE-BP model is established for the two basic problems of regression prediction and classification in the field of machine learning,and it is successfully applied to the actual production and life.(2)Aiming at the regression problem,SAE-BP soil temperature prediction model based on spectral features is established to predict the soil temperature in the target region pixels of the soil temperature dataset.In addition,the influence on prediction performance of the structure and parameters in SAE-BP model is also explored.The experimental results illustrate the effectiveness of the SAE-BP model in soil temperature prediction task.In addition,research on the adjustment and optimization of the structure and parameters in SAE-BP model shows that,within a limited range,increasing the depth of the hidden layers in stacked autoencoder and improving the labeled training data will improve the prediction accuracy of SAE-BP model to a greater extent.(3)Aiming at the classification problem,SAE-BP crop species recognition models based on spectral features,spatial features,and spacial-spectral joint features are established to identify the crop species in the target region pixels of the Indian Pines dataset.Multiple sets of comparative experiments are set up to verify the effectiveness of the SAE-BP model in features extraction and the influence on classification performance of the structure and parameters in SAE-BP model is also explored.The experimental results illustrate the effectiveness of the SAE-BP model in crop species recognition task.It is worth mentioning that hyperspectral data has the comprehensive characteristic of "image-spectrum merging".Our proposed SAE-BP model based on spacial-spectral joint features combines spectral information and spatial domain correlation characteristic of local homogeneity within the spatial neighborhood pixels,which greatly improves recognition accuracy.A grid search optimization on the parameters of the SAE-BP model is also performed,including the number of principal components,the size of the neighborhood window,and the depth of hidden layers in stacked autoencoder and BP neural network,and finally the optimal SAE-BP crop recognition model is obtained.In addition,when compared to the classical CNN deep learning models,the SAE-BP model shows stronger interpretability and universality.