<![CDATA[Teknik klasifikasi citra digital telah berkembang, dari berbasis pixel menjadi klasifikasi berbasis objek, dimana citra sebelumnya dibuat dalam bentuk segmentasi/poligon yang bias diatur homogenitasnya. Tetapi dalam proses klasifikasi baik dengan berbasis pixel dengan metode Maximum Likelihood maupun dengan berbasis objek tetap harus ditentukan training sampel untuk mengidentifikasi objek yang akan diklasifikasi. Dalam pengambilan training sampel dengan berbasis pixel, poligon yang dibuat, diambil sehomogen mungkin sedangkan dalam metode berbasis objek, training sampel dibuat berdasarkan poligon-poligon yang sudah terbentuk hasil segmentasi yang dibuat berdasarkan parameter scale, shape, compactness yang telah ditentukan. Penelitian ini bertujuan untuk menguji akurasi hasil training sampel yang dibuat berdasarkan polygon hasil segmentasi dengan training sampel yang dibuat berbasis pixel dengan studi kasus kawasan hutan di PLG Kapuas, Kalimantan Tengah dan citra yang digunakan citra Landsat. Akurasi diuji dengan melihat percampuran antar kelas (dengan Scatterplot) dan keterpisahan antar kelas dengan metode Confusion Matrix (nilai overall accuracy dan nilai kappa). Hasil memperlihatkan bahwa uji keakuratan training sampel berbasis objek pada lokasi lebih rendah ini jika dibandingkan dengan training sampel berbasis pixel, terlihat dari nilai Overall Accuracy dan nilai Kappanya. Grafik Scatterplot menunjukkan masih ada ketercampuran antar kelas (hutan, non hutan, non vegetasi dan tubuh air) pada kedua hasil dan lebih banyak terjadi pada training sampel hasil segmentasi.Kata kunci: training sampel, uji keakuratan, segmentasi, klasifikasi berbasis objek dan pixel, hutan dan non hutan, citra Landsat.ABSTRACTDigital image classification techniques have been developed from a pixel-based to an object-based classification, where the previous image is created in the form of segmentation/polygons whose homogenity can be set based on scale, shape, and compactness. However, in the classification process, either using pixel-based or object-based, several training samples still need to be determined in advance to identify objects that will be classified. In the pixel-based, while generating training samples, created polygons were made as homogeneous as possible. On the other hand, in the object-based method, training samples were made based on polygons from the results from segmentation process based on scale, shape, and compactness parameter. The research aim is to test the accuracy of training samples from the object-based method, which is compared with the ones from the pixel-based method. As the case study was forest areas around PLG Kapuas, Central Kalimantan. Landsat imagery was used as material. The accuracy was tested by looking at the values of inter-class mixture (using scatterplot) and of class-separation (using confusion matrix to gain overall accuracy and kappa value). The results show that the accuracy of pixel-based training samples is better, which can be seen from the Kappa value and Overall AccuracyScatterplot graphic shows that there are mixed-classes (forest, non-forest, non-vegetation, and water bodies) on both samples test result, although there are more in the segmentation process rather than in the training samples made from manual delineationKey words: training samples, test accuracy, segmentation, object and pixel-based classification, forest and non forest, Landsat imagery]]>
