Research On Land Use Cover Classification Optimization Methods For Hyperspectral Remote Sensing Image Based On ELM

The use of remote sensing(RS)images for land use coverage classification is an important part of obtaining land cover information,and it is also the current focus of land use and land cover change(LUCC)research.In recent years,due to the great advantages of hyperspectral remote sensing images,such as multi-bands,high resolution,and rich information,using them to classify land use has become a research hotspot in the field of remote sensing.However,the size and high-dimensional features of hyperspectral remote sensing data have also brought great challenges to the research of remote sensing image classification.Traditional classification methods,when applied to the classification of hyperspectral remote sensing data,are likely to result in excessive processing time,computational complexity,and local minimum problems.Such methods are difficult to meet the needs of current classification applications with hyperspectral remote sensing images,especially in terms of classification performance and efficiency.Extreme learning machine(ELM)is a rapid machine learning algorithm proposed in 2006.It exhibits the advantages of fast learning,high efficiency,and good generalization performance when dealing with large-scale data.Therefore,applying the ELM algorithm to hyperspectral remote sensing image classification can effectively overcome the development bottleneck in this field.This paper applies ELM to the classification of hyperspectral remote sensing images.It mainly focuses on the problems of instability,poor robustness,and low classification accuracy of ELM remote sensing image classification methods.From the three perspectives of ensemble learning,making full use of image texture features and deep learning,three kinds of ELM-based remote sensing image classification optimization methods are proposed in this paper.The specific research content is divided into the following parts:(1)In addressing the instability of the classification results in the ELM classification,combined with ensemble learning ideas,an ELM remote sensing image classification method based on ensemble learning is designed and implemented.First,multiple training sets are generated by resampling,and every ELM-based classifier is trained separately;then the base classifiers with unstable or poor classification results are deleted,and the base classifiers are integrated by using the combination of the voting method and maximum probability method.Finally,this method was used to classify the Indian Pines and PaviaU hyperspectral remote sensing data.The results show that this method can not only enhance the stability and robustness of remote sensing image classification while ensuring the efficiency of ELM classification,but also improve the classification accuracy.(2)For the low classification accuracy of ELM,considering the rich spatial texture features of hyperspectral remote sensing images,a KELM remote sensing image classification method based on Local binary pattern(LBP)texture features was proposed.Firstly,this method uses minimum noise fraction rotation(MNF)transform to reduce the dimension of the band,then extracts the rich texture features of the remote sensing image by the LBP operator,and then uses the radial basis function as the kernel function to construct the KELM classifier.Finally,it is applied to the land cover classification.The results show that LBP-KELM has achieved good results in the classification application,with high classification accuracy and less time consumption.Compared with the classification map of remote sensing images that is based on the method proposed in(1),the ELM classification map is smoother and the pitting points are significantly reduced.(3)Combining the cutting-edge theory of deep learning,in order to make full use of the respective advantages of the deep learning algorithm and the ELM algorithm,a hyperspectral remote sensing image classification model based on convolutional neural network(CNN),CNN-ELM was designed and implemented.This method uses CNN's convolutional layer and sub-sampling layer alternately to construct the depth feature extraction layer,and uses ELM to construct the classification layer,then it is implemented with Keras,a popular deep learning developing framework.Finally,both the effectiveness of the classification model and its advantages in classification accuracy are demonstrated with experiments.In summary,the effectiveness of the three optimized ELM-based remote sensing image classification methods proposed in this paper have verified with the experiments.While ensuring the advantages of ELM algorithm in terms of classification efficiency and speed,these optimized methods further improve the accuracy of land cover classification of hyperspectral remote sensing images,and enhance the stability and robustness of the basic ELM algorithm.