Science Service System
You are here : Home : Proposals_Summary

Summary of Proposal LAN0599

TitleMapping and monitoring of the protected areas of the Bwindi National Park using TerraSAR-X data
Investigator Otukei, John Richard - Makerere University, Department of Geomatics and Land Management
Team Member
Prof. Blaschke, Thomas - Salzburg, Geography and Geology
Prof. Dr Collins, Michael - University of Calgary, Department of Geomatics Engineering
Prof.Dr. Strobl, Josef - University of Salzburg,
SummaryUp-to- date information on land cover and land cover changes is critical for protected forest area management. The traditional approach for land cover and land cover change information extraction is through the laborious, time consuming and often subjective analogue methods, notably, photo interpretation and field work measurements. Over the past decades, data from satellites is increasingly being used for mapping and monitoring of physical environments and process thereof. There are many satellite sensors onboard remote sensing platforms providing data with varying spatial, spectral, temporal and radiometric resolutions. Data from these sensors can be classified as either optical or microwave depending on region of operation within the electromagnetic spectrum. Both optical and microwave sensors have specific problems that limit there applications. It is therefore envisaged that a combination of optical data with high spectral resolution such as Landsat TM with microwave data with high spatial resolution will offer benefits especially for forest cover mapping in the tropics. The overall aim of my study is to critically examine the potential of remote sensing data, both optical and microwave for land cover mapping and monitoring of the protected area of the Bwindi National Park. The scope of the study covers both the horizontal structure (Land cover) and vertical structure (Biomass). First, the research assesses the performance of Synthetic Aperture Radar, Landsat TM and IKONOS data for Land cover mapping in the Bwindi National Park. An Object Based Image Analysis approach is used for image segmentation prior to image classification. Alas, image segmentation results in many object features that can be used for developing rule sets for classification. This challenge is addressed using a semi automated approach integrating Object Based Analysis (OBIA) and Decision Trees (DTs). Secondly, the research investigates the utility of SAR and Landsat data for biomass estimation. Knowledge of biomass helps not only to know how much a given ecological region acts s as a potential for carbon sink or source but also as a means for forest monitoring. Decreasing biomass content will imply forest degradation as well as increased carbon emissions while increasing biomass content signifies forest re-growth and decreased carbon emissions. Unfortunately, the estimation of the biomass and/or potential carbon emissions especially in the tropical forests with mixed vegetation is known to be a difficult task. Conventionally, the estimation of biomass can be done through: 1) regression analysis, 2) mean tree method, 3) and unit area method. These approaches although useful do not provide accurate biomass estimates since the final estimates have to rely on the results from the sample sites. The challenge therefore lies on the development of cost effective techniques for estimating the potential carbon emissions in a reliable and reproducible way. The emerging technologies in the field of Geomatics are envisaged as a key and perhaps the most viable way of establishing the carbon estimates. Consequently, the study utilizes the data provided by Landsat and TerraSAR satellite systems for biomass estimation. The study in particular examines the potential of radar backscatter for biomass estimation. The analysis is performed in combination with ground based field measurements i.e. tree height, diameter at breast height and crown cover. The ground based measurements are used as proxies for estimating the above ground biomass. It is envisaged that the results obtained will enhance the potential for future decision making regarding the conservation of the forest ecosystem. The Phd program is partly funded by Austrian Agency for International development
Final Report1. Pixel based analysis of TSX data One objective of the study was to analyse the potential of increased polarisation for improved land cover classification. In order to assess the potential of additional polarisations, a priori classification based on single TSX polarisations was performed. The classification using the HH polarisation using Decision Trees (DTs) provided an overall kappa index 0.4323. Using the same training and test data, a classification of VV polarisation, provided an overall kappa index of 0.3577. A further classification based on the two combined polarisations resulted in an overall kappa index of 0.4657. Three observations can be deduced from the analysis of the dual TSX data: first, neither the HH nor the VV polarisation classified all the selected classes of interest in this study and that each polarisation has varying detection rates for selected land cover types. Secondly, a combination of HH/VV showed an increased potential for land cover mapping with an overall kappa index of 0.4657 which was higher than the values obtained using independent analysis of either HH or VV polarisations. Third, observation was that, while analysis of combined HH/VV polarisations improved the classification accuracy, overall, the resulting accuracy was low for detailed land cover mapping. The low classification accuracy was caused by interclass confusion resulting in high omission and commission errors. Interclass confusion can be attributed to several factors such as the high correlation between the HH/VV polarisations providing similar intensity values in the respective classes leading to low separability as well as the effect of layover, foreshortening and shadows. Analysis of TSX data was also performed using supervised Wishart classifier resulting in overall accuracy of 57.45% and corresponding average accuracy of 43.9%. 2. Addition of derived bands This study also examined potential of derived SAR texture for improved land cover mapping. Both pixel and object based methods were adopted. Three types of textures were considered for pixel analysis: The SAR specific textures (SARTEX), textures based on grey level co-occurrence matrix (GLCM) and textures based on image histogram (HISTEX). Analysis of the HH/VV original polarisations with a single SARTEX (VI) resulted in a kappa index of 0.6740 which represents a 45% increase in classification accuracy. A similar analysis using GLCM and HISTEX resulted in kappa indices of 0.6655 and 0.7106 respectively. Accordingly using the GLCM and HISTEX increased the classification accuracy by 43% and 53% respectively. However, unlike the SARTEX where only one texture measure was used (VI), 3 texture measures were used for both GLMC and HISTEX. Nevertheless the resulting accuracies using SARTEX, GLCM and HISTEX had comparable accuracies. A classification using single GLMC and HISTEX measures were performed but the resulting accuracies were less than that obtained using the VI texture. A further classification using a combination of VI, GLCM and HISTEX resulted in an increased classification accuracy represented by a kappa index of 0.7535. Increased classification accuracy can be obtained through analysis of both original TSX and derived texture measures. By and large, a single SARTEX performed better than single textures in either GLCM or HISTEX. While this study has demonstrated the potential of SAR texture for land cover mapping using TSX data, it would have been desirable to compare the results with existing studies. Unfortunately, there are no available studies that have examined TSX textures for mapping in tropical environment. The SAR texture analysis was also performed in the context of OBIA. The bands used for analysis of TSX data, included the two HH/VV original TSX bands, the SAR specific texture (VI) as well as the standard deviation (SD), dissimilarity (DIS), correlation (CORR), entropy (ENT), angular second moment (ASM), mean (MN), contrast (CON) and homogeneity (HOM). Through feature space optimisation, the MN, CORR, VI, HOM, ENT, ASM, VV, and HH combination provided the highest average separability and hence were selected as the most appropriate features for classification. Tests results showed overall classification accuracies of 64.84%. High class confusion was obtained with most of the classes but visual analysis indicated that the image segments resulting from multi-resolution segmentation were representative for the classes including those that were poorly classified. 3. Image fusion Image fusion was also carried out using both pixel and OBIA methods. For pixel approach high pass frequency filtering (HPF), principal component analysis (PCA) and wavelet principle component analysis (WPCA) were used resulting in overall accuracies of 71.86%, 82.78% and 83.60% respectively with corresponding overall kappa indices of 0.6876, 0.8082 and 0.817. After post classification filtering using a 3 by 3 majority filter, the classification accuracies increased to 74.99%, 83.12% and 85.38 with corresponding kappa indices of 0.7220, 0.8100 and 0.8369. A classification using original Landsat ETM+ provided overall classification accuracies of 92.41% and 98.32% before and after the application the majority filter with overall kappa indices of 0.9148 and 0.9811 respectively. It is evident from the results that none of the image fusion methods between the original Landsat ETM+ and HH polarisation of TSX image provided improved accuracy in comparison to the classification obtained using the original Landsat ETM+. Also, of the three selected image fusion methods, the WPCA provided the highest overall classification accuracy. Furthermore, out of the three image fusion techniques, the PCA had colour distortions while the WPCA and HPF preserved the original colours of the Landsat ETM+. A further image fusion referred to as eCofusion was carried out using the OBIA approach. Prior to eCofusion, object based classification was performed using Landsat ETM+ data resulting in overall classification accuracy of 84.13% with a corresponding kappa index of 0.822. For the selected image fusion techniques, eCofusion 1 (excluding zeros in TSX) had the highest classification result of 87.31% while eCofusion 2 (including zeros in TSX) and eCofusion 3 (excluding TSX in classification) had overall accuracies of 84.21% and 83.56 % respectively. Through a direct comparison with the results obtained using pixel based methods, the following observations were made: 1) Object based method provided low land cover classification accuracy obtained using Landsat ETM+ data compared with the accuracies obtained using pixel based methods 2) unlike the case of pixel based methods, it was possible to obtain a high classification accuracy through multisensor data fusion using OBIA. Again the reduced classification accuracy with other image fusion techniques implemented using OBIA were caused by effects of relief distortion on TSX data. This effect was more pronounced with the results obtained using pre-image fusion with wavelet transform.

Back to list of proposals

DLR 2004-2016