Research Of Remote Sensing Image Classification Methods Based On Distance

The number of obtainable remote sensing images has dramatically increased due to the rapid development of science and technology,and the improvement of remote sensing observation techniques.Large amounts of remote sensing data contain a wealth of useful information which playing an important role in social services,economic development and national security.At present,the applications of remote sensing data have extended to the military,research and civil fields.It is a difficult task to obtain and use the available information effective in the specific applications.Traditional image processing methods cannot be applied to the processing and information extraction of remote sensing images directly because the unique characteristics of remote sensing images.We need to exploit the remote sensing image processing technology.Remote sensing image processing technology can process and analysis remote sensing images automatically by using pattern recognition and machine learning techniques to extract the useful features.Remote sensing image processing still unable to meet the needs of various users though it has been made a lot of progress in visual,digital,networked and intelligent.As the growing demands of remote sensing data applications have accelerated the development of remote sensing image processing.An emphasis has been placed on automatically processing and analyzing these remote sensing images and extracting useful information from them.Remote sensing image classification is a method used to automatically extract features and distinguish distinct objects in images.Remote sensing image classification is one of the key technologies of remote sensing image processing.Remote sensing images classification plays an important role in government decision making and public services.Many other fields also require remote sensing image classification techniques.Although scholars have make a lot of research about remote sensing image classification and presents a series of related methods,but existing accuracy of remote sensing image classification was still unable to meet the needs of specific applications.This dissertation studies the unsupervised classification and context classifier for remote sensing image.Unsupervised classification of remote sensing images without much prior knowledge and classify based on the spectral characteristics of the pixels on the image simply.It is simple to implement and with highly practical in applications.But the performance of remote sensing image unsupervised classification methods is not ideal,particularly when the difference between objects in spectral features is small.Contextual classifiers of remote sensing images taking into account the unique spatial characteristics in remote sensing data.Contextual classifiers are more accurate and work better compared with the traditional classifiers.But research of the contextual classifiers start late,there is plenty of room for improvement.Remote sensing image classification is the core content in this dissertation;we introduce the research status,basic principle and key technology of remote sensing image classification.And this dissertation focuses on two distance based classification methods.Including k-nearest neighbours(k-NN)and optimum-path forest(OPF).The experiment results illustrate the effective of our improvement for OPF and k-NN.The main contents and the results are as follows:We apply a new probability density function to the unsupervised OPF method.The used probability density function consider the cluster centres are surround by neighbours have low local density,and the distance between centres and samples have high local density is relatively large.That is,the cluster centres are characterized based on their densities and the distances between the centres and samples with higher densities.The proposed algorithm is used to segment remote sensing images.The experimental results demonstrate that our algorithm is superior to the original OPF method.We utilize the segmentation results of statistical region merging(SRM)method to construct the hybrid distance function and take full advantage of the spatial characteristic of remote sensing image.The merging predicts of SRM segmentation method is based on the adjacent regions,thus the segmentation result of SRM method reflect the spatial relation among samples of image.We can obtain different scales segmentation of an image by SRM,which can be expressed as a hierarchical structure.We compute the spatial distance function between pixels by the hierarchical structure of SRM results.Finally,we propose a hybrid distance function by combine the spatial distance with the traditional spectral distance.We apply the proposed hybrid distance function to the two classifiers based on distance: OPF and k-NN.We propose two contextual classifiers: k-NN-SRM and OPF-SRM.The traditional k-NN and OPF only use the spectral characteristic of remote sensing data and only depending on the spectral distance during the classification,which ignore the spatial characteristic of remote sensing data.We apply the hybrid spatial and spectral distance to these two classifiers based on distance.Therefore,we constructed two contextual classifiers: OPF-SRM and k-NN-SRM.The proposed two contextual classifiers consider both spectral and spatial characteristic during the classification.The experimental results on four real land cover images demonstrate these two contextual classifiers have good performance,which indirect proof the validity of the proposed hybrid distance function.