ANALISIS KETELITIAN KLASIFIKASI PENUTUPAN LAHAN MENGGUNAKAN METODE DIGITIZE ON SCREEN DAN DEEP LEARNING SERIES CONVOLUTIONAL NEURAL NETWORK (CNN) BERDASARKAN CITRA LANDSAT-8 OLI (STUDI KASUS: PROVINSI KALIMANTAN TIMUR)

Abstract

Land cover is dynamic due to human needs or natural events that can occur in a planned or unplanned manner. Dit. IPSDH-KLHK uses remote sensing satellite imagery data to generate land cover data using a visual interpretation method (manual interpretation). Object identification is done by digitizing on screen. Along with the development of the times and current technology, several studies have emerged regarding the classification of land cover and its accuracy test using the latest technology, one of which uses deep learning. In this study, the accuracy of the digitize on screen classification results and land cover classification using deep learning was carried out along with the accuracy test. The test for the accuracy of land cover classification as a result of digitizing on screen was carried out using the centroid method. Validation was carried out using high resolution satellite imagery, namely google earth pro according to the temporal acquisition of Landsat 8 imagery used, namely July 2019-June 2020 by spreading 360 samples randomly. The results show that East Kalimantan Province has 21 land cover classes with an overall accuracy value of 87.22% in the very good category and in the tolerance category. Land cover classification using deep learning is carried out using segmented Landsat-8 OLI images. Sampling was carried out with a segment picker for 20 land cover classes without the Mixed Dry Land Agriculture class. The classification results show overlapping because one land cover class is also classified into other classes and not all image areas are classified. The accuracy test was carried out with the same location of the sample point as the test sample for the digitize on screen method. The accuracy value of the deep learning method using 188 samples in classified areas resulted in an accuracy of 70.21% for 21 land cover classes. This is due to the many land cover classes with almost similar interpretation keys. The interpretation key of 21 land cover classes is more suitable for the digitize on screen method.