| dc.description.abstract | Most researches on the classification ofcoral reef benthic habitat mapping were conducted on pixel base analyses. This technique may lead to some extent of misclassification of coral reef ecosystem due to a high diversity and high spatial heterogenity. An alternative technique to reduce misclassification, an object-based image classification (OBIA) algorithm was proposed. This techniques first decomposed an image into relatively homogeneous objects or areas and then classified these objects (areas) instead of pixels. OBIA has been well developed and applied in terrestrial studies in past decades, but the method has not been used adequately on benthic habitat mapping. In addition to an image classification technique as one of the component in the analysis of remote sensing, the selection of a classification algorithm was also an important factor because it can largely affect the final result. Benthic habitat mapping studies of coral reefs are generally applied with traditional classification algorithms such as Maximum Likelihood (ML), which requires the spectral response of each class following the pattern of normal distribution (Gaussian distribution). In contrast to a contemporary algoritm, the machine learning algorithm has received more attention in the benthic habitat mapping for producing higher accuracy thanML classification algorithms. Currently, the machine learning algorithm has been widely used in the commercial remote sensing data processing both on pixel-based and object-based. The algorithm is also potential to be developed for coral reef ecosystem mapping. Application of an object-based classification technique with machine learning algorithm is still limited in Indonesia, especially for mapping of coral reefs, therefore, it requires an intensive study to develop this algorithm as an alternative for traditional algorithm. The general objective of this research was to identify the effectiveness of the application of satellite image classification techniques in generating detail mapping information of coral reef ecosystems. To achieve this objective, the following special objectives were developed: (1) to develop a systematic approach in the coral reef ecosystem classification to define the classification scheme, (2) to mapping coral reef ecosystems using object-based classification techniques with advanced classification algorithms (machine learning) on satellite image data, and (3) to assess status of coral reef ecosystems within the last few decades. Results of this study were expected to be used for consideration in the sustainable management of coral reefs ecosystem in the District of Morotai Island, especially in North Maluku, and can be used as a baseline to support the development of the field of mapping and remote sensing of coral reef ecosystems. This research was conducted in the coral reef ecosystem located in the western part of the Morotai island in North Maluku province. The study stepsconsisted of equipment preparation, data collection, data processing, and preparation of dissertation begins in September 2012. Field data collection was held in October 2012, while processing and data analyses were conducted from January v to September 2013. The method used in the study consisted of field data collection using the technique of line transects photo quadrat, pre-processing of Landsat 5TM, 7EMT + and 8OLI, the development of a classification scheme with ecological approach quantitative analysis based grouping in a hierarchical (Agglomerative Hierarhical clustering, AHC), image classification pixel-based techniques, and also developed object-based classification techniques with traditional classification algorithms and advanced algorithms (Support Vector Machine (SVM)), accuracy assessment of coral reefs thematic maps generated from Landsat imagery, and change detection habitat of coral reef ecosystems. The observation showed that the coral reef ecosystem on the Morotai island consisted of life coral, associations organisms, benthic vegetation, and abiotic substrates. Benthic components of coral reef substrate were dominated by sand and seagrass material. Classification schemes were developed using data from the percentage of cover components of reef benthic habitat and were capable of producing 10 classes. From the 10 classes, only 7 classes that can be used for classification using satellite imagery. The use of cluster analysis and similarity value had good ability to defines the classification scheme. The optimum value of both types of kernel SVM algorithm generated the overall accuracy of the number of whole classes that can be mapped to either the value C 8192 for linear kernel with accuracy value of 71% orwith the higher accuracy of the RBF kernel with a value of C 4096 accuracy value of 81%. The RDF kernel was a recommended kernel type of coral reef habitat for classification by the number of intermediate classes. The coral habitats can be better mapped using SVM algorithm except sand rubble. Object based classification techniques can improve the results of basic reef habitat classification of 18-25% better than pixel-based classification techniques. SVM classification algorithm can improve mapping accuracy better than the traditional classification algorithms on pixel-based classification techniques. Change detection analyses in this study were able to reveal dynamics changed of coral reef habitats that were applied to the satellite image classification techniques with minimal field data availability. Change detection technique showed the difference between the image of the 1996, 2002, and 2013 both quantitatively and qualitatively. Thematic maps change detection combined with statistical calculation not only identify broad classes of habitat change between the time of analysis, but also able to identify habitat changes. Variations of the coral reef habitat changes in the period 1996-2013 may be due to an extreme climate change and an increasing intensity of human activities on the coral reef ecosystem. | id |
