| Monitoring forest resources in China is being characterized as large scale and short cycle. The traditional methods of monitoring forest resources based on field inventory is becoming inadequate in terms of meeting needs and requirements of current and future forestry production and ecological construction. Remote sensing based forest resource monitoring technologies have become more and more popular. The overall objective of forest resource monitoring is to obtain forest classification maps and stand parameters including tree species composition, stand age, diameter at breast height(DBH), tree height(TH), tree crown width(CW), live crown height(LC), basal area(BA), leaf area index(LAI), etc. Currently, the research focuses on using satellite remote sensing imagery and estimating canopy closure, stand volume and biomass. But, there are few reports on modeling of TH, DBH, BA, etc., using remote sensing. In addition, there is a lack of comparing different methods and integration and use of multi-source data to monitor forest resources.Objectives of this dissertation are to develop remote sensing based methods to generate spatially explicit estimates of important forest biophysical parameters includng LAI, TH, DBH and BA using sample plot data for Chinese fir plantations at regional and national scale and thus to provide potential technologies for extraction and estimation of stand parameters at large scales. A state-owned Huang-Feng-Qiao Forest Farm in You Xian County, Hunan Province, was chosen as the study area. Terrestrial laser scanning data, SPOT5 and Worldview-2 high resolution images were used. The images were first pre-processed and then combined with field plot observations using process based and regression models. LAI, TH, DBH and BA were modeled at various scales. The obtained results included:(1) Study on selection of important factors for development of stand parameter modelsAfter atmospheric correction and ortho-rectification, spectral and texture characteristics of the remotely sensed images were analyzed. A total of 73 spectral variables including band reflectance, texture measures, vegetation indices, and other image derived spectral variables such simple band ratios(SR), moisture stree index(MSI), difference vegetation indices(DVI), and red gree vegetation indiex(RGVI) were extracted. Pearson product moment correlations between the spectral variables and stand parameters(LAI, mean TH, mean DBH and BA) were then calculated and the important spectral variables were selected based on significance levels of 0.01 and 0.05. Moreover, variance expansion factor(VIF) was employed to analyze multicollinearity between each other of the selected spectral variables. The partial correlations of the spectral varables with the stand parameters were further analyzed and the spectral variables that had VIF value greater than 10 were eliminated. With the above selection procedure, a total of 18, 13, 14, and 11 important spectral variables were obtained for LAI, mean TH, mean DBH and BA, respectively.(2) SPOT5 image based stand parameter modelingWith the selected variables above, stand parameter models for mean TH, mean DBH and BA and LAI were developed using stepwise regression, partial least squares regression and nonlinear regression. The results and accuracies of the obtained models were evaluated and compared using coefficient of determination(R2), standard deviation of the residuals, model evaluation index value(MEI) and residual scatter plots. Results showed that the optimal models for mean TH and mean BA were obtained using partial least squares regression with R2 values of 0.653 and 0.774, respectively, whereas the optimal models for LAI and mean DBH were yielded using nonlinear regression with R2 values of of 0.773 and 0.676, respectively. It was found that the spectral variables that were frequently involved in the models were SR12, a modified band ratio RSR, DVI34, Red band, SR and SR21. The results indicated that mean TH, mean DBH and BA and LAI could be successfully retrieved using SPOT5 satellite images.(3) Worldview-2 image based tree crown and stand parameter modelingThe information of tree crown width(CW) for Chinese fir within each of sample plots was extracted from Worldview-2 image using marker watershed algorithm and mean shift segmentation. With different parameters, the results from these two algorithms were compared. The results showed that both methods accurately led to extraction of tree crowns, but the performance of the multi-scale mean shift segmentation was much better than the marker watershed segmentation. For the marker watershed segmentation, the accuracy of 77.1% was obtained with the minimum segmentation threshold of 10. The accuracy of the multi-scale mean shift segmentation was 86.6% with the segmentation parameters of 10 sh =, 6rh = and M =20. In terms of visual interpretation, the segmentation resulted in excellent match of the tree crown boundaries. However, the multi-scale mean shift segmentation led to better match than the marker watershed segmentation. Meanwhile, a novel method was proposed to improve the smoothing of tree crown borders using a probability median value. This method led to a good match of the smoothed boundaries with the tree crown borders in images and eliminated the jagged and gourd-shape features. With the smoothed tree crown values and TH and DBH measurements, the nonlinear simultaneous equations of DBH and TH with tree crown were obtained and the resulting R2 values were 0.899 and 0.913, respectively, indicating the models were reasonable.(4) Terrestrial 3D laser scanning based stand parameter retrievalSingle trees and stands were scanned using a terrestrial 3D laser scanner and point cloud data of trees and stands were obatined. In order to extract information of DBH, TH and CW, the point cloud data of single trees were first analyzed. The analyses included segmentation and reduction of point coud data using database technology and noise in the cloud data was removed based on statistical methods. The location, DBH, TH, and crown diameter of each tree were determined based on the frequencies of point cloud data at vertical direction along the X and Y directions. An algorithm to determine the location of each tree was proposed. This algorithm first divided, indexed and saved the point cloud data into uniform grids and the center coordinate of each grid was calculated. The location of each tree was then determined by judging the distances among the adjacent grids. Based on its location, the point cloud data of each tree within a sample plot was derived and its DBH, TH and CW were then extracted. The results showed that the estimates of DBH and TH had significant linear relationships with its filed measurements with the R2 values of 0.987 and 0.883, respectively.Overall, this study led to following three innovations: 1) A novel method to improve the smoothing of tree crowns was proposed based on a probability median value based algorithm. After smoothing using this algorithm, the jagged features due to the raster data of the extracted tree crowns were eliminated and the boundaries of the smoothed crowns matched very well with their original shapes; 2) Nonlinear simultaneous equations to estimate TH and DBH for Chinese fir plantation using the smoothed crown widths were developed; and 3) A novel algorithm to determine the location of a tree in a stand and to obtain its point cloud data from terrestrial 3D laser scanning was proposed. |
PDF Full Text Request