Research On Classification Using Land-Use Image Based On High Performance Computing

Automatic recognition technology of remote sensing (RS) images computer is always arousing more attention in RS research field. Classification method research for RS image is an important part of this field, which directly affects the application range and effect of RS data. In recent years, image classification technology is developing to get more intelligence classification technology, to dig deep image information, to classify combined with multi-intelligence classifier. It leads to the result that high spatial resolution image data widespread increases, imaging spectral band becomes narrow and geometry and texture information of surface features become more noticeable.In order to increase classification accuracy, we can utilize texture and shape features to remove "salt and pepper noise"effect. But computer has overload to process image data and hardly meet actual requirement as the reason of RS data high-resolution need and geometric progression computation, so one new image classification method through parallel computing tends to arise.The purpose of this paper is to research the classification of high-resolution RS image, analyses and sum-ups pixel-based classification and object-based classification respectively for RS image, combines with image characters and assistant space knowledge in order to improve the accuracy of classification of RS image. Especially for the features of mass data amount for high resolution image and of large and complicated computing for multi-scale segmentation and classification algorithm, this paper designs one image classification module based on high-performance geological environment which utilize the advantage of some parallel calculation architectures to improve classification accuracy and calculation efficiency.Currently, main problems existing in high-resolution image classification are:(1) Pixel-based image classification can not fully utilize high-resolution image features, bring about the difficulty to obtain classification accuracy from image itself.(2) Object-based classification accuracy depends on image classification result to great extend. Segmented object easily create deviation and these deviations accumulate through every select and classification process, and finally directly influence integral classification result and this effect is irreversible.(3) For either classification method, one hand, they can improve classification accuracy during calculating mixed with many aspects of spectrum, texture, shape and context for high-resolution image data. On the other hand, with accompanying geometric progression computation, we have to decrease dimension, decrease computing parameters or something like that to improve efficiency, but in the meantime, decrease classification accuracy definitely.This paper introduce one parallel calculation module which combines pixel-based and object-based image classification methods together and is designed as the goal of high-resolution image classification based on high-performance geological environment. The main research aspects are concluded as below.(1)Data capture becomes quicker, more intelligently and updated frequently. The depth and the breadth of application field is developing. It analyses the bottleneck problem of large scale high-resolution image data processing requirement, borrows from rapidly growing parallel calculation technology and network technology(including distributed computing, Grid computing, cloud computing networking and etc.), proposes a HPC-RSCM application module of high-resolution image classification under the high-performance geological calculation system in order to resolve complicated calculation problem of mass data calculation field.(2)It analyses two classification methods, that are pixel-based and object-based classification, sum-ups each advantages and disadvantages. For pixel-based classification method, it analyses ACC(Ant Colony Clustering), fuzzy clustering, SVM(Support Vector Machine) good points and bad points and also improvement ways, proposes an intelligent combination classification method of multiple classifiers in order to improve image classification accuracy and efficiency. For object-based classification method, it analyses error in classification which possibly arise in classification process(multi-scale segmentation, characteristics optimization, fusion algorithm classification), proposes improved fusion algorithm method, PSO-ACO optimization in order to decrease error an increase integral classification accuracy.(3)In the practical application, we wish that detail spatial constructurc and distribution information of terrestrial object can automatically transfer to available information among mass data to service for each field need, meanwhile we find such clear problem as bigger and bigger data, complex background information, serious noise interference, synonyms spectrum, foreign body in the same spectrum. These practical problems expand high-performance calculation technology application in image data processing. This paper analyses the technology and proposes two operable parallel classification modules under the support of multiple classifiers fusion module. that are simple parallel classification module combined by multiple classifiers and complex parallel classification module combined by multiple classifiers. On the thought of proposed HPC-RSCM module, it designs an image classification system under different construct parallel environment which can classify and practice in MPI parallel construct, MapReduce cloud architecture, GPU common calculation architecture and OpenMP parallel module of SMP system respectively.(4)Test analysis. Through testing SPOT5image data and comparatively analysing the results of segmentation, characteristics optimization and sort algorithm, meanwhile comparatively analysing parallel efficiency of computing under high-performance different construct environment, this paper testifies that these two parallel classification modules are feasible in practice both improving image classification accuracy and handling mass image data. The final result is to obtain realistic high-performance geonomy computing architecture for mass data image classification and also find a feasible way to transparent access large scale computing resource sharing.Next task is to research, resolve and improve remaining problems that is how to apply image data classification method to fully improve computing power of parallel platform, and how to realize automatical movement of platform with classification method.