University of Twente Student Theses


Impact of topography and species diversity on the prediction of forest metrics from VHR multispectral imagery

Hossain, Md Sarowar (2022) Impact of topography and species diversity on the prediction of forest metrics from VHR multispectral imagery.

[img] PDF
Abstract:Multispectral remote sensing has been extensively used for estimation and monitoring of forest structural attributes. However, consideration of the influence of landscape factors such as topography and species diversity in forest attribute estimation are comparatively limited. To date, very few studies were found that evaluated the accuracy improvement of forest structure estimation models by incorporating topographic influence in the model and even fewer studies were found that investigated the changes in prediction accuracy by the influence of coniferous species diversity especially with multispectral imagery. Those studies that included topographic variables in their models, did not conclude how the topography effect the relations of other explanatory variables with forest attributes and whether or not that addition improved the model accuracy. The aim of this study was to investigate: 1) how strongly the texture variables derived from VHR multispectral imagery correlate with forest metrics, which are mean diameter at breast height (DBH), standard deviation of diameter at breast height (SD DBH), and tree count per plot. 2) The changes in relationship between texture variables and forest metrics by the influence of slope and aspect and therefore any increase the models’ estimation accuracy. 3) Any further improvement in models’ estimation accuracy if the data is separated based on species diversity of the study area. This study utilized the World View-2 derived texture variables calculated with different parameter settings to assess the relationship with field measured forest metrics. An iterative subsampling procedure was followed to fit stepwise regression models for each forest metric with the significantly correlated texture variables to determine the most significant variables and develop the prediction model, while the subsampling approach minimizes the spatial autocorrelation issue. Then, different models were developed adding slope, aspect, and their combined influence as interaction term in the stepwise prediction model running the subsampling algorithm and compared in terms of R2, RMSE and AICc. Moreover, the best fitted models were used to predict forest metrics in different species diversity forests. The correlation coefficients of significant texture variables for Mean DBH ranged from -0.47 to -0.57, 0.36 to 0.41 for SD DBH and -0.51 to -0.56 for Tree Count. Interaction effect of slope and aspect on texture variables significantly changed the relationship with forest metrics in most cases and slope as a moderating variable, improved models’ R2 by 15%, 6%, and 11%, and the RMSE was decreased by 1.03, 0.54, and 0.3 for Mean DBH, SD DBH, and Tree Count, respectively. Aspect influenced model showed an increased R2 by 4%, 5%, 5% and decreased RMSE by 0.29, 0.47, 0.17 for Mean DBH, SD DBH, and Tree Count, respectively. The best fitted models predicted Mean DBH, SD DBH and Tree count with an R2 of 0.54, 0.45, and 0.42 and RMSE of 6.03, 3.86, and 3.73, respectively. Moreover, the splitting of model dataset based on species diversity showed that Mean DBH and Tree Count was predicted in single species forest stands with average R2 of 0.60 and 0.50 and average RMSE of 5.70 and 3.48, respectively and SD DBH was be predicted in multiple species forest stands with average R2 of 0.68 and average RMSE of 2.80. The correlation of texture variables with forest metrics was highly sensitive to GLCM parameter selection used to calculate the textures. The relationship of texture variables with forest measured variables changes significantly when texture variables have interaction effect of topographic variable. Therefore, forest structural attribute estimation accuracy can be improved in mountainous region by incorporating influence of topography and prediction can be even more improved if the model is fitted with species specific data.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Subject:38 earth sciences
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page