Research And Application Of The Object-oriented Segmentation And Classification Based On Spot5Remote Sensing Image

This paper used the SPOT5remote sensing images of Xiaomei region of Zhejiang provinceas data sources and the eCogniton software as operating platform，conducted a research ofsegmentation and classification experiment and application study which was based on object-oriented. Through the evaluation of different segmentation scheme， the optimal segmentationresults as boundary extraction was obtained; The objects after segmentation were classified and theattributes were extracted.By comparing with the classification results of traditional maximumlikelihood，the conclusion that object-oriented classification can achieve higher classificationaccuracy was obtained; The dynamic change and change trend of forestland was analyzedaccording to the results of two phase image classification;Based on the data of forest resourcesmanagement survey in2009， the forestland boundary lines were modified according to thetechnical regulations for defining forestland borders. Finally the results of boundary and classattribute extraction were generated.Analysis the status of forest land protection and utilizationbased on the results of forest land border definition and make the planning. Details are as follows:(1) The image preprocessing mainly includes Geometric correction，Image cutting andImage fusion. The fusion methods of HIS，PCA,Gram-schmidta and Pan were used.The resultswere evaluated indexs of Standard deviation， Information entropy， Average gradient，Correlation coefficient， Deviation index and Spectral distortion.By synthesizing the visualevaluation and calculation results，confirm the Pan is the best fusion method in the study area.(2) The multi-scale segmentation experiment was conducted in the eCognition software withdifferent segmentation schemes. On the basis of visual assessment， choose the results of scale80,90and100to take the objective evaluation. Based on the existing category of forest sub-compartment boundaries and combined with the visual interpretation to determine the referenceobject.According to the rule of UMA and Similarity criteria， calculated the objects with thefollowing indicators which are Roundness， Compactness， Shpae Index， Circumferencerelative error， Area relative error and Absolute displacement of center.Then got the conclusionthat when the Band weight is1:1:1:1， Shape factor is0.2， Compactness factor is0.5，Segmentation scale is80， the egmentation result is most suitable for the forest sub-compartmentboundary extraction in the study area.By manually modifing， it can be used as a fieldinvestigation work hand chart.(3)The objects were classified after segmentation.According to the actual situation of thestudy area， we divided the forest land as bamboo forestland and other forestland.The non-forestlands were divided as water，agricultural land,construction land and unused land. The finalclassification accuracy is90.80%with the kappa coefficient is0.88. Compared with the traditionalmaximum likelihood classification results，the overall accuracy increased5.8%and the kappa coefficient increased0.08. So we choose the object-oriented classification method to classify thetwo phase images.The two phase images classification results were applied to the study ofdynamic change of forest land. The statistical results show that:The forest land area is decreased，reduced41.89ha. The bamboo forestland increased28.94ha and the other forestland area reduced70.83ha.The reduction of forestland is mainly transfer to farmland and unused land.(4) On the basis of the datas of forest resources management survey in2009in Xiaomeiregion， according to the technical regulations， the forestland boundary lines were modifyiedusing the results of boundary and class attributes extraction as reference.Based on the results of themodified forest boundary lines,we compile the forestland protection and utilization planning aferanalyzing the situation of forest in Xiaomei region.