Pengembangan Metode Analisis Citra
INTEGRATED MODEL OF WATER COLUMN CORRECTION TECHNIQUE FOR IMPROVING SATELLITE-BASED BENTHIC HABITAT MAPPING: A Case Study on Part of Karimunjawa Islands, Indonesia
(MSc Thesis at the Masters Program for Watershed and Coastal Management)
Pramaditya Wicaksono (08/279233/PGE/0768)
Under supervision of Prof. Hartono and Projo Danoedoro
One of the most important factors that must be understood on the management of coastal area is the distribution of benthic habitat. Benthic habitat is an economically and ecologically important natural resource in coastal area. The distribution of benthic habitat can be well-presented using map. Benthic habitat map is a powerful tool for coastal management planning such as locating protected area, prediction species occurrence, evaluation of management effect and also biodiversity assessment. One of the cost-effective methods to provide such information with fast, high repetition and accurate result is using satellite image to map those resources in combination with field observation. Remote sensing technology allows us to produce good habitat maps. The problem is thehabitats are located submerged and limit the ability of remote sensing data to map the habitat.
The aims of this study is to inverse the submerged reflectance into wet reflectance using integration of concepts on how pixels value over benthic habitat is recorded by sensor, water column attenuation, and bathymetry data. The accuracy of the integrated model will be compared with the accuracy of the existing model on each classification scheme. Last purpose is to combine the integrated model with PCA to get more detailed information on benthic habitat types. The hierarchical benthic habitat classifications were derived from ecological basis and habitat spectral separability analysis. The purpose of using hierarchical scheme is to cover all the possible existing habitat and to cover different management needs. The methods used in the study were conversion into surface reflectance, sunglint removal, water column and bathymetry generation, water depth invariant index, PCA (Principle Component Analysis) transformation and integrated model. Digital classification process was carried out using maximum likelihood with knowledge-based image segmentation.
The result shows that the integrated model could inverse the submerged reflectance into wet reflectance but produced slightly lower accuracy compared to Lyzenga or PCA on each habitat classification scheme due to discrete zoning in bathymetry data. In classifying 5, 7, and 13 habitat class, Lyzenga was the most accurate with 79.59%, 74.73% and 37.89% accuracy respectively, PCA produced 82.65%, 71.57% and 37.89% respectively, and the integrated model produced 73.46%, 69.47% and 37.89% accuracy respectively. The combination of integrated model with PCA produced the most accurate result on the detailed classification scheme with average accuracy of 61.56% and overall accuracy 50%. The integrated model competes well with other methods on each classification scheme, especially on detailed classification scheme.
Key words: Remote sensing, coastal management, benthic habitat, hierarchical
classification scheme, water column, integrated model.
AIRBORNE HYPERSPECTRAL DATA FOR MANGROVE SPECIES MAPPING
A comparison of pixel-based and object-based classification
(MGIS thesis at School of Geography, Planning and Architecture, The University of Queensland, Australia)
Under supervision of: Professor Stuart Phinn
Mangroves are one of the most important objects of wetland ecosystems, which are threatened by both anthropogenic and natural disturbances. It is necessary to monitor and assess these environments to assist conservation and restoration efforts. Remote sensing methods are required to effectively and accurately mapping and monitoring mangroves due to their location and extent. The advance in remote sensing technology, especially the hyperspectral sensor, has significantly improved its ability to differentiate earth surface objects into higher detail, including mangrove species. This study aims to investigate the ability of high-resolution hyperspectral data, CASI-2, for mangrove species mapping using a pixel-based approach and object-based approach, and examine the different results between them.
The study area is located at the mouth of the Brisbane River area, southeast Queensland, Australia. The primary dataset used in this study was CASI-2 hyperspectral dataset with 4 m spatial resolution and 30 optimised bands for vegetation studies. A comparison of three classification methods with different analysis domains has been performed, which were spectral angle mapper (per-pixel), linear spectral unmixing (sub-pixel), and a combination of rule-based and nearest neighbour classification (object-based). The existing mangrove map from Queensland Herbarium/EPA has been used as guidance for the selection of targeted mapping objects and as reference for evaluating resulting maps. In conjunction with the pixel purity index (PPI) map, nine endmembers have been collected as the targeted classification for pixel-based approaches, namely closed Avicennia, closed Ceriops, closed Rhizophora, open Avicennia, shallow saltmarsh, medium saltmarsh, deep saltmarsh, vegetated saltmarsh, and water body. A combination rule-based and sampling method was incorporated to identify these classes in the object-based approach.
The spectral angle mapper (SAM) compared the angle between the endmembers' vector and each pixel vector in n-dimensional space. The pixel with the smallest angle with the certain endmember was assigned to the corresponding class. The resulting map showed solid class boundaries with only few unclassified pixels with overall accuracy of 69% (Kappa 0.57). Linear spectral unmixing (LSU) accounted for the abundance values of each endmember for every pixel. The results were images describing the endmembers fraction abundance on each pixel. The final classification map showed a patchy pattern of classes and there were many unclassified objects with overall accuracy of 56% (Kappa 0.41). Object-based (OB) approach analysed the image in object space rather than in pixel space, which consists of sequences of segmentation and classification procedures (together form the rule sets). The resulting map has very solid polygons of classes with overall accuracy of 76% (Kappa 0.67).
Post-classification comparison (PCC) has been employed to quantify the degree of class agreement between reference map and resulting maps with five classes, and among resulting maps with nine classes. In comparison to the reference, SAM has the highest degree of class agreement, followed by OB and LSU. The agreement percentages were 60.4%, 57.3%, and 41.2%, respectively. Among resulting maps, SAM-LSU has 47.3%, SAM-OB has 57.7%, and LSU-OB has 42.7% of classes match. The mismatches were mostly found within closed Avicennia and saltmarsh classes, this was due to the highly mixed pixel in these classes. Other factors influencing the low results were the division of saltmarsh into finer detail (for among resulting maps comparison) and the coarse resolution of reference map. This study indicated that SAM, which fully regarded the endmembers and worked on a pixel basis, was the best classifier to map the distribution of mangrove species. However, object-based approach was also proven to be very effective in delineating mangrove species patches.
OPTIMALISASI KLASIFIKASI DIGITAL TUTUPAN LAHAN DARI CITRA PENGINDERAAN JAUH MENGGUNAKAN ALGORITMA BAYESIAN MODEL AVERAGING CLASSIFIER
Pembimbing: Hartono dan Projo Danoedoro
Penelitian ini didasarkan pada asumsi bahwa dengan heterogenitas dan kompleksitas kenampakan permukaan bumi, khususnya objek tutupan lahan, diperlukan algoritma klasifikasi parametrik yang lebih fleksibel terhadap pola spektral tutupan lahan. Adapun tujuan penelitian ini adalah melakukan kajian optimalisasi klasifikasi digital tutupan lahan dari citra penginderaan jauh dengan cara 1) memodifikasi dan mengimplementasikan algoritma Bayesian Model Averaging Classifier (BMAC) dengan berbagai model distribusi data untuk klasifikasi multispektral tutupan lahan dari citra penginderaan jauh; 2) menguji akurasi hasil klasifikasi modifikasi algoritma BMAC dengan berbagai kombinasi model distribusi data kemudian membandingkannya dengan akurasi hasil klasifikasi algoritma Gaussian Maximum Likelihood Classifier (GMLC); dan 3) menguji efisiensi waktu eksekusi algoritma BMAC dengan berbagai kombinasi model distribusi data.
Algoritma BMAC diimplementasikan menggunakan Interactive Data Language (IDL) agar menjadi sebuah perangkat terintegrasi dengan perangkat lunak pengolah citra digital. Lebih dari 500 baris kode program ditulis untuk mentransformasi algoritma BMAC menjadi sebuah perangkat lunak antarmuka grafis interaktif. Beberapa model dan parameter statistik tertentu yang digunakan dalam BMAC mengalami simplifikasi dan modifikasi sebelum proses implementasi, untuk tujuan efisiensi kode dan waktu eksekusi program. Hasil implementasi algoritma BMAC diujicoba pada Citra Landsat ETM+ multispektral, dengan 15 kelas tutupan lahan. Akurasi hasil klasifikasi BMAC kemudian dibandingkan dengan akurasi GMLC menggunakan metode Confusion Matrix. Selanjutnya efisiensi waktu eksekusi algoritma juga diuji dengan menggunakan satuan waktu eksekusi relatif terhadap waktu eksekusi GMLC.
Hasil klasifikasi menunjukkan bahwa secara keseluruhan akurasi GMLC lebih akurat dari BMAC. Dengan demikian, hipotesis bahwa akurasi keseluruhan hasil klasifikasi digital tutupan lahan akan meningkat dengan kombinasi model distribusi pada algoritma BMAC, ternyata tidak terbukti berdasarkan hasil penelitian ini. Akan tetapi, hasil klasifikasi juga menunjukkan terdapat beberapa kelas tutupan lahan yang memiliki Producer’s Accuracy atau User’s Accuracy yang lebih tinggi pada BMAC. Kenyataan ini merupakan sebuah indikasi bahwa tidak semua kelas memiliki nilai spektral terdistribusi normal, sehingga asumsi GMLC tidak selamanya benar. Terkait uji efisiensi, eksekusi algoritma BMAC berjalan jauh lebih lambat dibanding GMLC. Karena bentuk persamaan BMAC yang lebih kompleks dari pada GMLC, selain itu juga terdapat beberapa algoritma ekstensi dalam BMAC seperti Bayesian Information Criterion, dan berbagai aktifitas teknis seperti kalibrasi nilai diskriminan. Hasil uji akurasi dan uji efisiensi algoritma menunjukkan bahwa, GMLC merupakan algoritma yang paling optimal untuk klasifikasi digital tutupan lahan. Dengan demikian, tujuan optimalisasi klasifikasi digital tutupan lahan menggunakan algoritma BMAC, secara umum tidak tercapai dalam penelitian ini. Akan tetapi, dengan ditemukannya beberapa kelas yang memiliki akurasi individu lebih tinggi pada model-model non-Gaussian, menunjukkan bahwa BMAC optimal untuk kelas-kelas tertentu.
Kata kunci: Maximum Likelihood, Bayesian Classifier, tutupan lahan, klasifikasi digital
LINEAR SPECTRAL MIXTURE ANALYSIS FOR STUDY LAND COVER CHANGE AT URBAN REGION USING LANDSAT TEMPORAL SATELLITE DATA (Case Study in Banjarbaru City)
Supervisors: Projo Danoedoro and Sutikno
As viewed from remote sensing perspective, urban region is a veru heterogeneous area, which gives reflectance from various land cover types and materials. The limitation of the spatial resolution from moderate resolution sensors like Landsat is that it requires analyses at sub-pixel level. Mixed pixel in remote sensing data is one of the sources of error in accuracy assessment in conventional classification. This research tried to apply Linear Spectral Mixture Analysis (LSMA) method to detect land cover change (vegetation, impervious surface, bare soil and water) at sub-pixel level in Banjarbaru City based on Landsat multitemporal data.
LSMA is an approach incorporating analysis of sub-pixel, which can give information on the fraction in each pixel, so that it offers a potential solution for pixel-based classification problems. In this study, a Maximum Likelihood Classifier was applied as a comparsion for the LSMA. Accuracy assessment of this method was based on the higher spatial resolution of IKONOS image. Several processing stages were applied in this research to increase the accuracy, i.e. Atmospheric Correction, Minimum Noise Fraction (MNF) and Pixel Purity Index (PPI). Change Detection was done with multitemporal image differencing method.
The percentage of each land cover component in each pixel was performed by the fraction images, which derived from the LSMA method. By this model, with the error average is 0.017 for 1992 image and 0.016 for 2003 image, which indicate that each endmember of land cover has been associated well with small deviation standard. The accuracy assessment of abundance for each endmember by using IKONOS image showed a value of 95%, which indicates that LSMA provides a high accuracy result to detect the endmember of land cover at sub-pixel level.
Change detection result showed that the vegetation land cover in the study area increased by 2.84% for the entire study area, bare soil increased by 30.42%, and water body increased by 20.2%, while the impervious surface decreased by 35,31%. The change detection result found that the LSMA can show the change within each pixel or in every area at size of 30x30 m, which is presented in terms of increase and decrease of endmember fractions. This information was found very useful to compensate the lack of such ability performed by the Maximum Likelihood algorithm.
Keyword: MNF, PPI, Linear Spectral Mixture Analysis, Maximum Likelihood, Change Detection