Research On Hyperspectral Remote Sensing Image Feature Extraction Based On Improved Filtering Algorithm

Hyperspectral remote sensing image(HRSI)can provide tens or even hundreds of bands of information rich in the earth's surface.From these information,the effective extraction of feature information is of great significance to the research of many related fields,such as ground object recognition,scene understanding,target detection and so on.However,HRSI provides abundant useful waveband information for data analysis,but also brings many problems and challenges for hyperspectral remote sensing image analysis and processing:(1)Noise problem.In the process of acquiring HRSI,a lot of complex noises often exist in the image due to the influence of sensors,atmosphere and illumination changes.This will seriously distort the band information,which is not conducive to extract feature information effectively.(2)Small sample problem.It is difficult and costly to acquire real object labels from HRSI,so it is not easy to collect a large number of label samples for training models.This leads to the problem that only a small number of labeled samples can be used to classify HRSI,which is called small sample problem.(3)Cross-region mixing problem.In the process of denoising HRSI,cross-regional mixing often occurs because of the small spatial resolution and the complex distribution of the objects,that is,there are other features besides the target features.If the denoising task is performed in this case,the output image will be blurred due to cross-region mixing,which will affect the feature extraction of HRSI.In this paper,the key techniques of robust feature extraction and classification for hyperspectral remote sensing images are studied.Innovative achievements have been made in the research of noise problem processing,small sample problem processing and cross-region mixing problem processing for HRSI:(1)Processing algorithm of HRSI noise by using bilateral filtering technology based on classification and optimization.In the traditional bilateral filtering(BF)process,the template used in the denoising process will assign a larger weight to those dissimilar-structure pixels with near spatial distance,which reduces the denoising effect and make against the feature extraction of HRSI.Therefore,an improved bilateral filtering algorithm based on classification and optimization was proposed to solve the problem that the denoising effect of bilateral filtering is not ideal.By classifying and optimizing in the BF template,the algorithm selected pixels with similar structure to generate a new template,eliminating the influence of structure-dissimilar pixels on filtering denoising.This ensures that the new pixels have better features in weight distribution,and improves the denoising effect of HRSI,so as to improve the effect of feature extraction.The experimental results show that the method is effective in denoising and simple and effective.(2)Processing algorithm of HRSI noise based on trilateral smoothing filterWhen traditional BF is used to denoise HRSI,if the central pixel of the neighborhood is just a noise point,the BF can not extract characteristics of the HRSI very well.Because the spatial proximity measure function in the BF algorithm is insensitive to noise,and when the neighborhood center is a noise point,the gray similarity measure function can not well express the actual similarity between pixels.To compensate for these shortcomings,this paper constructed a neighborhood mean similarity judgment function to improve the existing BF algorithm.The results show that the feature extraction method based on trilateral smoothing filter has a strong denoising ability,which can make the extracted features retain effective spatial and spectral features,and improve the classification accuracy of HRSI classifier.(3)Processing algorithm of HRSI Small sample based on superpixel bilateral filtering.When using BF to denoise HRSI,there are two aspects worth considering.On the one hand,if the distance between the unstructured similar pixels and the target pixels is relatively close,the output value may be greatly affected,which will limit the effect of BF on the weighting restriction of unstructured similar pixels.At the same time,although BF restricts the dissimilar structure pixels by weighting,it still distributes the weight of the non-structural pixels,which still have an impact on the output value.On the other hand,the characteristics of hyperspectral images are different from those of general images for existing many homogeneous regions.Pixels in those regions are more likely to be structurally similar,and characteristics of the target pixel can be enhanced by making use of domain-structurally similar pixels.According to the limitations of BF and the characteristics of HRSI,if theose homogeneous regions of hyperspectral image can be segmented reasonably,and then they are filtered by BF respectively,the limitation of BF on unstructured similar pixels will be greatly improved.In this way,the extraction of image features will be more obvious and more separability.Based on BF,a superpixel BF feature extraction algorithm was proposed in this paper.The algorithm combines superpixel and BF to extract spatial and spectral features of HRSI in depth,and obtains more obvious features.Even though small samples retain the original dimension,they can get high classification accuracy.Experimental results show that the proposed method has made a breakthrough in small sample classification and improved the classification performance of the classifier.(4)Processing algorithm of HRSI cross-region mixing based on propagation filtering.Due to the limitation of small spatial resolution and the complexity of terrain distribution,cross-region mixing often occurs in the denoising process of HRSI.In this paper,propagating filtering technology was introduced.When cross-region mixing occurs,the cross-region mixing pixels can be avoided or effectively alleviated by not assigning or assigning as little weight as possible.The results show that the feature extraction method based on propagation filtering can effectively solve the cross-region mixing problem and improve the classification accuracy of the classifier.