Object-oriented Classification Of Forest Cover Using SPOT5 Imagery

The classification of forest cover is the most important and essential task in employing remote sensing to forest resources inventory and monitoring.Obvious advancement had been achieved in theory and technology during past decades,thought the classification can still not full met the need of practical application,for the classification accuracy need to be further improved in the southern China,where were with fragmentized distribution of forest,species and types diversity,and complex structure.With the high spatial resolution remote sensing data are available recent years,remote sensing can be expected to applicate wildely in forest resources monitoring,and object-oriented methodology also provides us a hope method to image analysis.However,the existing theory and technology of remote sensing had been built on low or medium resolution imagery and based on pixel,so it is important and urgent to keep on study to imagery process,information extract,imagery classification for the high or very high resolution imagery and meaning object.Aiming toward improve the classification accuracy of SPOT5 imagery,a synthetical and systemic research was focused on the image processing,image segmentation,employment of Landsat 7 ETM+ image as ancillary,spectral and texture features extraction and selection,forest cover classification with multi-classifiers and combination,the main works and result are showed following:1)To collect the ground control points(GCPs)for the geometric correction of the hi-spatial resolution remote sensing imagery and employment in forest resources inventory and planning,a GPS control network was found in the study area.The test results indicated that,the RMS error of single plot of GCPs positioned by WADGPS was lower than 0.5 m, which completely met the demands of geometric correction of the high-spatial resolution remote sensing image,such as SPOT5,QuickBird and IKONOS.Using GPS control network as the reference station and solving the high-precision coordinate conversion parameters,the largest RMS error of single plot positioned by portable GPS was 3.86 m.The average boundary displacement and absolute displacement of sub-compartment's center were 3.23 m and 3.76 m respectively and the accuracy of area survey was higher than 98%,which full satisfy the practice implication.The accuracy and efficiency of the application of GPS in forest district was improved effectively after establishing the GPS control network.2)SPOT5 data processing methods for forest resources inventory in south china were studied.It was indicated that fusing data using IHS transformations and resampling by bilinear interpolation took much advantages in preserving spatial,texture and spectral information of image than any other methods,linear stretching by segment and edge enhancement were helpful to reveal forest information and improve classification accuracy.3)The strategy of imagery segmentation and information extraction combining SPOT5 HRG imagery and Landsat 7 ETM+ imagery was studied,and the selection method of object features was discussed.ETM+ data couldn't take part in the imagery segmentation,for their coarser resolution would reduce the homogeneity within objects,lead to the objects contained impurities and couldn't represent the distribution of forest cover exactly.Spectrum and texture were most useful object features for the classification of forest cover in plantation.Most object features had been eliminated after selected by covariance,sample correlative coefficient and multiple correlative coefficient,the features remained were with the specific meaning,independence with each other,and had definite information.4)A five-step object-oriented classification routine was present to do classification on SPOT5 imagery,it included imagery segmenting,rule-based classifying,classification-based segmenting,area control,multi-classifier classifying and combining,topper layer synthesizing. Five classifier were employed to the classification,included minimum distance classifier, Mahalanobis distance classifier,Bayes rule classifier,fuzzy classifier and support vector machines.The result of classification indicated that in the study area with the fragmentized distribution forest,species and type diversity,complex structure,with the Bayes classifier,the total accuracies of the third level,the second level and the first level were 79.38%,81.82% and 86.98%respectively,where the third level contained twenty-two class based on age-group of trees,the second level contained fifteen class based on species,and the first level included nine class based on species groups.Bayes rule classifier also got the closely producer accuracies of all classes.With hierarchical classification,the result of beginning from the low layer and then synthesizing upward topper-layer was better than that classifying from top-layer to low layers.Used as ancillary data,Landsat 7 ETM+ data were helpful to improve the classification accuracy of SPOT5 imagery.5)The combination methods of multi-classifier were studied,a new combination method, defined as voting rule/fuzzy fusion,had been developed,which combined the conservative voting rule and fuzzy fusion.It got the better result than voting rule and fuzzy fusion through the test of combining five classifiers.The accuracy had been improved by combining classifiers,although the improvement was not as obvious as expected previously.