PhD Defence Mr Yaseen Taha Mustafa

Department of Earth Observation Science

Title of defence

Improving Forest Growth Estimation: Bayesian networks for integrating satellite images and process-based forest growth models

Summary

Carbon sequestration through forestry has the potential to play a significant role in ameliorating global environmental problems such as atmospheric accumulation of greenhouse gases (GHG) and climate change. Estimating the contribution of forests to carbon sequestration is commonly done by applying process-based forest growth models, such as the Physiological Principles Predicting Growth (3-PG) model. Output of such a model, however, suffers from uncertainties and requires a close attention of their inputs. Moreover, with the development of modern satellite remote sensing imagery, Moderate Resolution Imaging Spectroradiometer (MODIS) sensor also monitors the forest and its growth. Satellite images provide extensive data of forests with a large spatial coverage. These data, however, are not free of uncertainties and are often not available. To address the reduced uncertainty and limited data availability, this thesis presents an approach which has been constructed using graphical statistical models, in particular Bayesian networks. This approach was applied to the Speulderbos forest in The Netherlands, where detailed data were available. In order to explore and solve the issues and requirements of this approach, five studies were carried out in this thesis.

First, a Gaussian Bayesian network (GBN) was used to address the bias in a process-based forest growth model and the noise within satellite images. A novel inference strategy within the GBN was developed to take care of the different structures of the inaccuracies in the two data sources. The obtained outputs with the GBN were more accurate than either the 3-PG model or the MODIS estimate. Hence, the GBN improved the estimation of the forest growth estimates values by integrating a 3-PG model with MODIS imagery.
 

The Gaussian Bayesian network for ith iterations, i ≥1, every triple nodes LAIMi; LAIBNi and LAI3PGi indicates the network at time i and refer to MODIS LAI, GBN, and 3-PG mode, respectively.

Second, the performance of GBN modeling for improving forest growth estimates was evaluated. This was done by considering the quality of GBN output and relationships among GBN variables, in order to determine the reliability, the applicability and the robustness of the GBN. For that purpose, both evidence propagation, and a sensitivity analysis of the input sources, i.e. the 3-PG model output and MODIS observations were addressed. In the context of this study, a thorough assessment of the sensitivity of the proposed GBN aimed to achieve a better understanding of the potential of GBNs in modeling forest growth, and in particular to estimate the relevance of the GBN input parameters and the required level of precision needed to provide accurate estimates of forest growth estimates. Evidence propagation by means of the 3-PG model improved the GBN output, showing that its relative error with respect to the field data decreased by 2.0%. This improvement is stronger than propagating the same evidence through MODIS images as the relative error of GBN output increased by 4.5%. It was also found that a GBN is more sensitive to the variation of satellite estimates than to variation in forest growth model output. 

Sensitivity test of the Bayesian network after simultaneously varying theLAI_3PG and LAI_M values.

Third, the performance of GBN modeling for forest growth estimates was investigated and assessed with missing input sources, i.e. satellite data. The satellite time series, however, contained gaps caused by persistent clouds, and cloud contamination. The EM-algorithm was formulated and integrated within a GBN to estimate these missing values. During a period of 26 successive months, the EM-algorithm was applied by making synthetic gaps at four different cases: successively and not successively missing values during two different winter seasons, successively and not successively missing values during one spring season, and not successively missing values during the full study time. The estimated values well represented the original values where the maximum value of the averaged absolute error between the original values and those estimated was equal 0.16. Moreover, the root mean square error of the GBN output reduced from 1.57 to 1.49.

Fourth, GBN modeling for forest growth estimates with the EM-algorithm was applied to real gaps of satellite data. The Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) sensor was used as input source into the GBN. Moreover, the GBN was modified to improve forest growth estimates that show variation in space. The GBN output was more accurate than the 3-PG estimate, as the root mean square error reduced to 0.46, and the relative error to 5.86%. Moreover, the deviation of the averaged output of spatial GBN and field data was less than the deviation between the averaged output of the 3-PG and field data.

Finally, the spatio-temporal estimation of growth estimates of a heterogeneous forest using GBN was carried out. GBN combined the spatial version of the 3-PG model output with decomposed MODIS images. The Linear Mixture Model (LMM) was used to decompose MODIS pixels using class fraction derived from an aerial image and an ASTER image. In this way spatially heterogeneous output was produced. Results showed that the spatial output obtained with the GBN was more than 40% accurate than both the spatial 3-PG model output and the satellite estimate, with a root mean square error less than 0.53.
 
The study focused on improving forest growth estimates by using GBN to integrate two sources providing forest growth estimates, i.e. 3-PG model and satellite images. Such integration can be applied to different types of forests by giving more attention to remote sensing images. The study concluded that forest growth estimates could be improved using GBNs. Moreover, formulating the EM-algorithm within a GBN can adequately handle missing satellite data and improve estimation of forest growth estimates. Ultimately, this study contributed to provide the means for an accurate and reliable estimation of forest growth estimates.

Biography

Yaseen Taha Mustafa was born on 18 January 1979 in Mosul, Iraq. In 1996, he joined the Department of Mathematics, College of Computer Science and Mathematics, The University of Mosul, Iraq, where he received his bachelor degree in Mathematics. From 2000 to 2003, he was appointed as an assistant researcher in the Department of Mathematics, College of Science, The University of Duhok, Duhok, Kurdistan region, Iraq.

In 2005, he earned his MSc degree in Mathematics from the University of Duhok. From 2005 to 2008, he was appointed as a lecturer in the Department of Mathematics, College of Education, at the same University. In 2008, he began to pursue the present PhD research from the Department of Earth Observation Science, Faculty of Geo-Information Science and Earth Observation of the University of Twente (ITC), Enschede, The Netherlands.

His research interests include the area of remote sensing, including mathematical and statistical tools, such as Bayesian networks. During his PhD study, he received the Best Student Paper Award at the ASPRS conference 2011 in Milwaukee, Wisconsin, USA for his paper “improving forest growth estimates using a Bayesian network approach”.

  Full text  

previous page
more events