Object-based Analysis for Extraction of Dom inant Tree Species

Meiyun SHAO ，Xia JING，Lu WANG

1.College of Information Engineering，Tibet University for Nationalities，Xianyang 712082，China；2.College of Geometrics，Xi'an University of Science and Technology，Xi'an 710054，China

1 Introduction

Remote sensing and image interpretation technology have been widely used in large-area forest management in recent years.Based on the interpretation of image data，we improve our efficiency and save a lot cost.But in previous program，we considered TM，MODIS etc.Mid or low-resolution imagesmost for image interpretation process，form which we can only get the rough information of the forest like pure or not pure.We have no idea about detailed forest typeswhich can be very important in forest inventory program.With the high-resolution image occurring，it ismuch easier for us to getmore information from imagery data becausewe could seemore clearly even a separate tree.In spite of this，automatically classification of forest types is still problematic due to the phenomenon that different typesmay have the same spectral reflection characteristics.As the traditional classification method could not satisfy the demand the accuracy we need and the sault-pepper phenomenon of the result seriously affects our judgment.So，how to extract useful information from the high-resolution image data，and how to effectively classify the different types of forest deserves discussion.Fortunately，researchers explored a new try in image processing field.That is the object-based analysis.Previous studies on high-resolution image proved there ismuch information contained in spatial relations of pixels.The contribution of textural and structural information isalso very important in imageanalyzing process.With this method brought up，some researchers have been concentrated on taking advantage of the spatial information lied in the image.Chubey et al.（2006）got forest structure parameters using objected-based classification method.Herrera et al.（2004）identified the tree species in the non-forestarea.Shiba and Itaya（2006）estimated the forest volume in middle Japan based on high resolution images and take into account of environmental change.Lackner and Conway（2008）automatically divided the land cover using IKONOS image and the accuracy is up to 88%.Mallinis et al.（2008）tested the multi-resolution image segmentation on an area located in Greece and got a nice result.Bunting and Lucasextract tree crown information from mingled forest based on CASIdata and high-resolution data.Wang Changying et al.（2008）studied the land cover types of Yalutsangpo River district using SPOT-5 images combining NDVI and shape et al.texture parameter by object-based analysis.Generally，image object-based classification method includes four steps：image segmentation stage；object analysis stage；classification stage and the accuracy assessment stage.However，how to effectively use the object features and how to solve themixture of classes of each object.Wewill try to combine the two steps in a system.This paper explores a way to extract dominant tree species using the objectoriented analysis in eCognition combined with fuzzy classification means and present the possibility of automatically identify trees from high resolution images.

2 Main content of experiment

Object-based classification generally consists of three steps：（i）creation of image objects using an image segmentation algorithm；（ii）extraction of object features；（iii）classification using the features.This study we will try to integrate these steps to improve dominant tree species identification taking advantage of the highresolution image.

2.1 Image segmentationMulti-resolution segmentation is a region-merging algorithm which first proposed by Martin Baatz and Arno Schape.The whole process showed in Fig.1 is described as follows：at the beginning，pixels are firstly merged into many small objects or regions and then these small objects are continued beingmerged into larger regions.

This image segmentation requires several parameters decided by users according to demand.They are：（i）weights of associat-ed layers；（ii）a color／shape ratio closely related to spectral／shape criterion of homogeneity；（iii）a compactness／smoothness ratio according to the object shape；（iv）a scale parameterwhich decides how large the objects are.Heterogeneity in eCognition considers primarily color and shape of objects.The heterogeneity f is controlled by these parameterswe set.

whereΔhcolorandΔhshapeare the indexes of shape and color，respectively；ΔwcolorandΔwshapeare the weight of them.

In order to realizemulti-bands segmentation，we need to add another wcwhich defines the weight of all layers.

where nmergeis the number of pixels within merged objects；nobj＿2，nobj＿1，are the number of pixels before being merged in object 1 and 2 respectively；σcis the standard deviation within objects of layer c.

Shape heterogeneity describes the shape from the opposite two sides-the smoothness and compactness.

where l is perimeter of object；b is perimeter of object's bounding box.

2.2 Object featuresThe featureswe used in this study based on spectral and texture information calculated from the objects derived from segmentation.Besides spectral characteristics like mean value of each layer，vegetation index and stand deviation of each band，higher-order texture measurement such as GLCM（Grey Level Co-occurrenceMatrix）wasapplied in our study.GLCM is a tabulation of how often different combinations of grey levels occur at a specified distance and orientation in an image object.The character valuesof thismatrix are very useful in classification for presenting the DN change rule.

2.3 Fuzzy classification methodFuzzy classification is a classic soft classifierwhich takes some factors into account including uncertainty in sensor measurement，vague class descriptions and classmixtures due to limited resolution.Compared with crisp classification，thismethod change the＂true or false＂into the continuous range of［0，….，1］.Avoiding arbitrary sharp thresholds，fuzzy logic imitates the complexity of realworld much better than the simple boolean systems do.Fuzzy logic canmodel imprecise human thinking and can represent linguistic rules.Based on the fuzzy logicmembership function，we selectsome features to representing the dominant tree species in the study area.We used some experimental function to describe the change potential of different selected characteristics.In the classification process，different objects were labeled different degrees and at last decide the class considering comprehensible factors.

2.4 Study areaThe study was undertaken in amature forest ecosystem located in the Dailing distict of Yichun，China（Fig.1）.The study area is a part of Da Hinggan Ling.Forests in the study area consist mainly of coniferous species including Korea pine，larches and spruce occurring at the top of themountain or hillside.Deciduous forests consistingmainly of oaks and birch occurring in pure stands andmixed with conifers are presentat lower elevations.

2.5 Data preparationDigital image data was acquired over the study area by the AlOS satellite on 8 September 2010.This data set was consisted of single panchromatic band imagery with the spatial resolution of 2.5m and 4-band multi-spectral imagery with the spatial resolution of4m divided into the following spectral bands：blue（420－500nm），green（520－600nm），red（610－690nm）and near-infrared（760－890nm）.

2.6 Application of the proposed methodWe have two level segmentations.The first is to separate forest from other types of land cover.We use amuch bigger scale.Considering it is related to NDVI index，we add the NDVI band in the segmentation process.In this level，we use NDVIvalue as a condition to extract all the forestland.We set theweightsofevery band as1，the scale parameter is50 and shape 0.2，compact of object0.5.And then we use a smaller scale to segment the forestarea againmaking sure itwas divided into homogeneous objects.We set the weights of spectral bands as0.5，NDVI layer 1 and panchromic layer 1，the scale parameter is 30 and shape 0.1，compact of object0.7.The segmentation result is shown in Fig.3.

Having found that conifer and broad-leaved trees obviously different in reflecting near-infrared band，we separate them mostly under this condition.Comparing the sample data，pine and spruce have the detailed difference；we add the panchromic band information to separate these two kinds.As for oak and birch，we use the standard deviation of near-infrared band to identify them because they're so similar in spectral reflectance，sowe have to find some texture difference as additional information in this process.The whole process is expressed Fig.2 and segmentation result showed in Fig.3.We can get the classification result showed in Fig.4.We can see clearly in the image thatnon-forest area could be clearly seen in the red color though there are some shadow areaswere wrong classified.The dominant tree species were extracted.And we can seemost area covers the pine and this was the samewith the truth.And there were still some broad-leaved tree species could not be identified.

In this paper，we evaluated the classification resultby confusionmatrix showed in Table1.We gota high accuracy in classifying conifer forestwhile the oak and the birch were not that easily separated.Basically，we extracted the dominant tree species who occupied at least 65%space of sample area.

Table 1 Confusion matrix of result

3 Conclusions

We can classify most of the dominant tree species in the study area.Derived from the true sample，we can conclude that due to the differences of the spectral and texture information between each kind，we can separate conifer forest and broad leaved forest，spruce and conifer，and even oak and birch.We can identify the tree category if it takes 70%space of a sample area.Choosing some object as validations，we can get all the accuracy assessment indexes clearly in the confusion matrix.The overall accuracy is about87%and the kappa coefficient is 0.837.It reflects a high accuracy in this classification and it shows the potential to identify more detailed tree species using object-based analysis.Conclusion and FutureWork：From an objectwe can not only get the spectral information but also texture that will help a lot in the classification.It provides us somany characteristics to express each class andmake it easy to find some effective clues to extract what we need.However，our study also has some problems.（i）We can only identify the dominant tree species which occupied 65%of whole area.As to a more complicated situation，we still cannot solve.（ii）The accuracy assessment process is not that rigorous because we only have the point samples，for a further study we should use polygon samples.（iii）These thresholds used in the experimentwere settled based on the datawe use，as for other situations，theymay not fit.（iv）We have notmade any quantitative assessment of the segmentation result.Next，we can try to find a good way to assess it.