Research On Individual Tree Identification And Crown Segmentation Algorithm In High Spatial Resolution Remote Sensing Imagery

To date,the increasing volume and readily availability of high spatial resolution imagery provides an effective source to extract the individual tree information for detailed knowledge of forest stand at different spatial scales. Nevertheless,conventional pixel-based classification methods are far from the use of improved spatial resolution satisfactorily. Recent years ,object based image analysis method with various image engineering techniques becomes a strong support for high spatial resolution imagery object recognition. Automated or semi automated tree detection and crown delineation using high spatial resolution(pixel size is higher than 100 cm) remotely sensed imagery provides a potentially efficient means to acquire information needed for forest management decisions ,sustainable forest management ,mapping damage due to insects and disease and ecology research. Tree detection can provide estimates of tree abundance and spatial pattern that are useful for evaluating density and stocking objectives. Delineation of individual tree crowns can get crown diameter of tree to be used to model tree structural variables such as height,volume,or biomass. Conventional field based plot survey methods which using statistically based plot sampling designs use expert knowledge,but its procedure is highly subjective and highly labour intensive and costly. So automated tree detection and crown delineation with high spatial resolution imagery has the potential to provide the required information in a more objective manner,at lower cost,and with greater coverage than is attainable using field sampling.Image based automated tree detection and crown delineation algorithm has almost 15 year's research history in north America,north Europe,Germany and Australia. This thesis reviews the main remote sensing segmentation methods and diverse individual tree identification and crown delineation algorithms. We analyze the theoretic model,applicability,precision,experiment condition,verification method,error analyse and limitation of these methods. This thesis focuses on the development of two methods for individual tree identification and crown segmentation in the object based image analysis framework .The presented approach develops a tree top seeded based region growth tree detection and crown delineation algorithm for analysing QuickBird satellite images in Populus×xiaohei plantation even stand at Xue JiaZhuang wood farm in Shanxi Province of China. This algorithm use the crown model which is focus on basic radiometric properties of tree crowns. We develop a method in which vegetation classification and crown segmentation are derived under a unified framework. After multiresolution segmentation,we get image object segments for tree top seeds detection with NDVI and ratio NIR feature. Around theses seeds,we let them region growing in a cycle way. Some false seeds must be wiped off with given feature threshold. After quadtree segmentation for crown shape optimization,the same category region must be merged. Now we get a crown map of test area. We use 9 plots with different plantation density(crown closure) to validate the above method. Average tree numbers identification error is 18.9% ,R 2 = 0.4693.From comparing tree numbers of field work and software identification by tree matching ,the confusion matrix,overall accuracy,commission error,omission error is computed. The main result is this algorithm's accuracy,commission error,omission error far from crown closure. Computed crown diameters after program crown delineation has similar distribution of field measure crown diameters,but they have bigger values and more dispersed range. Through grouped plantation density results analyzing,we find the performance of this algorithm on 0.6 crown closure plots get well,omission error of 0.8 crown closure plots is high to 34%,commission error of 0.7 crown closure plots is high to 63%. Ultimately,our tree top seeded based region growth tree detection and crown delineation algorithm is an effective way to get segmented crown in real stand image. We suggest users choose suited features and parameter values try by try in forehand applying.The presented approach develops a new mathematical morphology based marker-controlled watershed crown segmentation algorithm for crown segmentation. This method is be put on the QuickBird satellite images in Populus I-72 plantation even stand at Nan Gen village Hai Kou town in Anhui Province of China. Segmentation using the watershed transform works better if you can identify or mark foreground objects and background locations. Our marker-controlled watershed crown segmentation algorithm follows basic procedure. First,a segmentation function is computed which put image's dark regions as the objects you are trying to segment and LOG edge detection is applied. Then,computing foreground treetop markers after morphological reconstruction by opening and closing for filtering. The connected blobs of pixels within each of the crown like objects are treated as foreground treetop markers. Third step is computing background markers for identifying crown area. The boundary of isolated crown or grouped crowns can be got after mathematical morphology distance transformation. Forth step is modifying the segmentation function so that it only has minima at the foreground tree top and background crown boundary marker locations. Fifth step is computing the watershed transform of the modified segmentation function. This algorithm does not take into account the classification and only gets the image segment for further analyzing. We overlap the segmentation result with original image by manually crown delineation. By visual appraise,this algorithm works well. Average tree numbers identification error is 32%.We discuss the improvement ways to get better results.In the object based image analysis framework,we discuss the best fit spatial resolution choosing for crown segmentation. We use the ratio of average crown diameter to pixel size for experience index weighing the fitting from literature. This ratio can explain some crown segmentation phenomena,such as why in some more fine spatial resolution image,the crown segmentation result is not good as expectation. We extend our mathematical morphology based marker-controlled watershed crown segmentation algorithm to aerial images with 0.25m spatial resolution in secondary planted conifer stand at Cu Lai mountain in Shan Dong Province of China. We overlap the segmentation result with original image by only simple visual appraising. The result is more crowns identified than the field work. The tree top foreground marker technique for this kind of stand must be improved. Using object based image analysis method,after crown area,shadow area and bare soil area classified,we put up a simple formula to computing the crown closure of the sample images. Average computer crown closure calculating precision is high to 80%.In the end of this thesis,conclusion and discussion is given. Two kinds of program methods for crown segmentation algorithms are put up. One method is using Component Object Model(COM) and the other one is using web service technology. We draw the outline of estimation tree level attributes in automatic interpretation method researched before and put it to automatic interpretation method of estimation of stand level attributes. For all,many aspects of raised tree detection and crown delineation algorithms must be improved for method's practicality,our research has contribution to the knowledge of this area. Some suggestions are in the end of this paper for tree detection and crown delineation by experience.The crown closure of forest stand and the choosen of best fit spatial resoulution image are the very important factors affect the result of the algorithms.