Feature Analysis And Collaborative Classification Of Wide-spectrum Optical Remote Sensing Images

Remote sensing technology can realize human’s long-distance and contactless detection of the earth.It plays an irreplaceable role in the fields of agriculture and forestry management,urban planning,resource exploration,and target detection.With the changing of the wavelength of the electromagnetic spectrum,the imaging mode of remote sensing image also changes constantly.With the wavelength from short to long,the imaging mode gradually changes from passive imaging of optical spectrum to active imaging of microwave spectrum.Even in the passive imaging spectrum,the imaging mode also gradually changes from the reflect ion imaging mode in the visible/near-infrared band to the radiation imaging mode in the long-wave infrared band(also known as the thermal infrared band).“Wide-spectrum” remote sensing image means that its band range spans a wide electromagnetic spectral band,so there are great differences in imaging methods between different spectral bands.For classification tasks,wide-spectrum optical remote sensing images not only provide rich multi-dimensional information in different spectral bands,but also bring many challenges.On the one hand,due to the differences in imaging methods,there are great differences in image characteristics between different spectral segments.The preprocessing methods suitable for reflection spectral segment images are often not well suitable for radiation spectral segment images,and it is more difficult to accurately extract multi-source and multi-dimensional features in wide spectral segments;On the other hand,compared with non wide band remote sensing images,the multi-source and multi-dimensional features of wide band images are more complex,so it is difficult to suppress redundant information and effectively select diagnostic features for each category,and thus can not get good collaborative classification accuracy.This paper takes the wide-spectrum remote sensing image covering visible band,near-infrared band,and the thermal infrared band as the core,and takes the wide-spectrum image collaborative classification as the goal.From the perspective of preprocessing,feature extraction and fusion,feature selection,and classifier design,this paper analyze and discusses the main problems faced by the wide-spectrum remote sensing image collaborative classification.Through in-depth study of the principle of electromagnetic radiation and imaging mechanism,The corresponding solutions are put forward,which is of great value and significance to the popularization and application of wide-spectrum remote sensing images.The problems and main research contents of this paper are summarized as follows:Firstly,the traditional methods do not fully consider the unique imaging characteristics of thermal infrared remote sensing images in the denoising process,and it is easy to transfer the normalized scaling error to the subsequent processing process.This paper proposes a denoising algorithm for thermal infrared remote sensing image.In this paper,the radiative transfer model is introduced into the restoration process of thermal infrared remote sensing image s.Based on Taylor’s bilateral estimation and the constraint of surface temperature,the resto ration of image emissivity layer information and the supplement of spatial information are realized.The experimental results show that compared with the traditional algorithm,this method can not only suppress all kinds of noise,but also effectively redu ce the subsequent transmission of scaling error,and has higher convergence speed and recovery performance,so as to lay a foundation for the further interpretation of the thermal infrared hyperspectral image.Secondly,based on the nonlinear correlation model between brightness temperature and emissivity,a feature inversion algorithm suitable for low emissivity targets is proposed.The traditional temperature emissivity separation algorithm is based on the linear assumption between brightness temperature and emissivity,and the inversion effect of low emissivity targets is not ideal.In this paper,the derivation of the nonlinear correlation model between brightness temperature and emissivity is realized,and it is proved that the linear assumption is only a special case of the nonlinear model under specific constraints.Based on this,a smooth temperature emissivity separation algorithm suitable for low emissivity targets is proposed.Experimental results show that the algorithm has wider applicability and can get better simultaneous interpreting results for different sensor types,different atmospheric conditions,and targets with different emissivity types.Thirdly,aiming at the heterogeneity of spatial features of different spectral segments caused by different imaging methods,an image fusion algorithm based on multi-resolution super-pixel low-rank expression and residual learning is proposed.The algorithm uses super-pixel blocks instead of traditional blocks as low-rank recovery units,and adaptively adjusts the spatial characteristics in the unit to maintain the stability of ground object types in the unit and suppress structural noise;At the same time,by constructing a guided linear filter,the fine spatial features of the reflected spectral segment image can be transferred to the thermal infrared image on the premise of protecting the spectral information of the thermal infrared image;Finally,the residual model established in the low-resolution layer is transferred to the high-resolution layer to realize the spatial super-resolution of the thermal infrared image on the premise of ensuring the spatial detail information.Finally,the fusion and enhancement of spatial features of wide spectral remote sensing images and the alignment of heterogeneous wide spectral images at the level of spatial features are realized.Experiments show that the thermal infrared image spatial fusion algorithm based on multi-resolution low rank guided filtering has a better spatial smoothing effect and edge sharpening effect for the thermal infrared remote sensing image.The super-resolution thermal infrared image has more fine spatial information,and can effectively protect the spectral information of the thermal infrared image for different region types.The algorithm not only protects the diagnostic properties of spectral features but also enhances the unity of heterogeneous spatial features,which is more conducive to improv ing the collaborative classification accuracy of wide spectral remote sensing images.Finally,aiming at the problem that it is difficult to maintain a balance between the complexity and diagnostics of multi-dimensional features of wide spectrum remote sensing images,combined with the deep learning theory,a classification algorithm of wide spectrum remote sensing images based on multi-dimensional feature adaptive selection is proposed.Firstly,the algorithm determines the similarity between categories by constructing the merging matrix,and then adaptively learns the best feature weight used to distinguish each similar category based on the neural attention mechanism,so as to realize the measurement and selection of many wide-spectrum multi-dimensional features,and further realize the fine classification of each similar category.Experiments show that the algorithm transforms the “one against others” classification problem into a “one against one”classification problem between multiple similar categories,which can effectively balance the information between categories,enhance the interpretabilit y of features,and obtain more accurate classification accuracy.