Developing a methodology for discriminating natural objects having spectrally similar features using very high resolution satellite imagery

Obtaining reliable and accurate information about the Earth's surface features in terms of their quantity, diversity and spatial distribution is essential for a wide range of applications. Remote sensing technologies offering an efficient tool for monitoring the Earth's surface in various spatial and spectral resolutions have long been used for gathering this valuable information. It provides powerful and cost-effective solution in particular to visualization and analysis of large study areas. The latest technological innovations in the design of remote sensing satellites and sensors offer new application opportunities in many fields including environmental monitoring and natural resource management. Satellite imageries acquired by these systems introduce more detailed information about the nature, properties and distribution of the Earth's surface objects. Unlike moderate-resolution satellite imageries, individual surface objects, such as trees and buildings are visible in high resolution imagery. This important feature provides opportunities to differentiate spectrally similar species and mapping individual surface objects. Image classification is the most preferred and effective method for the interpretation of satellite imageries. Thematic maps representing different characteristics of the Earth's surface are produced as a result of the classification process. The accuracy and reliability of the thematic maps have great importance for success of the applications. The use of high resolution imagery poses some important challenges in analyzing and interpreting of digital image data. One is that with higher spatial resolution data the percentage of mixed pixels is increased because of the spectral similarity between adjacent pixels. In addition, due to the high degree of within-class and between-class spectral similarity of land cover units, information extraction or classification process becomes more complex. For these reasons, traditional classification techniques have been reported to be insufficient for producing accurate thematic maps using high resolution images. Therefore, novel and efficient analysis techniques are required for processing and analyzing such images. The overall aim of this thesis is to develop a methodology for classifying natural objects having spectrally similar features using high resolution WorldView-2 (WV-2) satellite imagery. For this purpose, image classification stages including pre-processing, classification and thematic mapping were comprehensively investigated in order to determine the most effective classification model for the particular classification problem considered in this thesis study. The dissertation starts with introduction part including problem definition, the purpose of the research, followed by methodology and extensive literature review. In the second part, general information about the basic principles of remote sensing including electromagnetic spectrum and electromagnetic energy, energy interaction with the atmosphere and the Earth's surface materials and spectral signature of surface features were described. Some image preprocessing techniques, such as corrections for radiometric, atmospheric and geometric effects were also presented. In addition, image transformations and the most popular transformation techniques (i.e. vegetation indices, textural features, principal component analysis, and Hue-Saturation-Intensity transformation) were given. Moreover, the detailed information about the image classification including supervised and unsupervised techniques, traditional pixel-based and object-based approaches, advanced classification algorithms namely, support vector machines, decision trees, ensemble learning algorithms including boosting, bagging, multi-boosting, random forest and rotation forest were discussed. At the end of this part, feature selection techniques were summarized and the several filter and wrapper based feature selection algorithms were discussed. The application part of the thesis was discussed under three sub-sections. In the first sub-section (Section 3), WV-2 satellite imagery and related pre-processing operations (radiometric, atmospheric and geometric corrections), study area and field applications including spectral measurements with a field spectroradiometer and analysis of spectral features of measured tree types were explained in detail. In addition, several pansharpening algorithms (i.e. HSI, PCA, Brovey and Gram Schmidt) were assessed to investigate fusion of a high spatial resolution panchromatic image with lower spatial resolution multispectral images of WV-2. Besides the preprocessing of satellite imagery and field applications, preparation of auxiliary data sets (i.e. vegetation indices, HSI transformations, principal components and textural features) were presented. Finally, land use/land cover types and sampling regions, the main inputs for the classification process, were determined by means of the spectral measures and atmospherically corrected WV-2 imagery. The second sub-section of application part (Section 4) addresses the object based classification and related processes. This section consisted of image segmentation, describing image object features, creating training and testing data sets, feature selection and object based classification. Region-based scale parameter selection was suggested and applied to perform image segmentation. Moreover, for the selection of relevant features and defining optimum feature size, 13 feature selection algorithms were applied to the high dimensional data containing variety of spectral and textural features of image objects. Data sets of different sizes formed by the feature selection results were used for the object based classification. Also the traditional nearest neighbor algorithm and advanced classification algorithms were applied to these data sets and their classification performances were analyzed thoroughly. The final sub-section (Section 5) focuses on the traditional pixel based classification method. To make objective comparison, training and testing data sets were determined in accordance with the data sets considered in object based classification. Feature selection algorithms were employed to the high dimensional data to produce new data sets having smaller number of spectral features. Pixel based classification was performed using these data sets formed by the feature selection results and the performances of classification algorithms were analyzed in detail. The results and discussion part of the dissertation (Section 6) includes detailed and comparative analysis of the application results, such as comparison of classification approaches and algorithms in terms of their use and performances, analysis of the effective features and optimum feature set size, thematic map producing and exploring the map accuracy. All in all, some important conclusions can be drawn from the results produced in this dissertation. First, the application presented in this thesis demonstrated that the conventional classification approaches were not suitable for the classification of high resolution satellite imagery as they found to be insufficient for spectral separability of natural objects. When the classification performances were compared, it was found that with the use of object based classification approaches, improvements in thematic map accuracy reached to 8%. Secondly, the results noticeably indicated that support vector machine and rotation forest ensemble learning algorithms outperformed the traditional nearest neighbor classifier and the other learning algorithms in both pixel based and object based classifications. Therefore, these two outperforming algorithms were preferred classification algorithm to perform accurate object based classification. Thirdly, from the analysis of measured spectral characteristics of tree species considered in this thesis, it was revealed that deciduous and coniferous trees have distinct spectral reflectance especially in near infrared region. As a result of the integration of the field-based spectral measurements and satellite imagery it was observed that red edge, near infrared-1 and near infrared-2 bands of WV-2 imagery provides valuable spectral information to differentiate between deciduous and coniferous trees. However, due to the high degree of within-class spectral similarity, multispectral bands of WV-2 image were found to be limited for differentiate spectrally similar tree types (e.g. Cedar and Cupressus trees). For this reason, instead of using only multispectral bands of WV-2 imagery, auxiliary data such as vegetation indices, textural features and principal components should be taken into consideration to increase the accuracy of the thematic maps. When the classification results were analyzed it was observed that the use of auxiliary data led to a clear increase in classification accuracy up to 8%. Fourthly, results showed that a reduction of high dimensional data to a few but specifically selected features also improve the classification accuracies. For this reason, a suitable feature selection algorithm such as SVM-RFE should be considered when high dimensional data were considered. Another important finding was that the all classification algorithms were affected by the curse of dimensionality or Hughes phenomenon. Thus, determination of optimum feature set size was also critical for the success of classification.