Analisis Kehandalan Ekstraksi Garis Tepi Bangunan dari Data Foto Udara Menggunakan Pendekatan Deep Learning Berbasis Mask R-CNN

Abstract

: The need of large-scale base map, especially in 1:5,000, is increasing in Indonesia. Furthermore, as the Government of Indonesia has declared 1:5,000 RBI mapping acceleration as one of main priorities of One Map Policy implementation, the need of large-scale topographic map production is also rising. Generally, topographic map feature extraction, including building extraction, is conducted through digitization or manually through feature stereoplotting either from satellite imagery or aerial photography. However, this method is usually time-consuming especially for high building density area mapping. Detection and extraction of building footprint automatically using computer vision of optical imagery have been favoured in recent years due to the time effective process. One of the technologies that have been developed is deep learning approach. However, the building line resulted from deep learning has disadvantage, i.e., irregular building footprint. This study attempts to assess the accuracy of polygon regularization resulted from automatically extracted building footprint using Mask Region-base Convolutional Neural Networks (Mask R-CNN) from aerial photography. The study finds that in high building density area with regular roof shape (AoI 1), the intersection over union (IoU) index is 87.8%. Whereas in high building density area with irregular roof shape (AoI 2) has the IoU index of 82.6%. This study also assesses the positional accuracy of 25 building corner point samples and resulting CE90 of 1.183 m and 1.303 m in AoI 1 and AoI 2 respectively. The geometric horizontal accuracy is classified as the class 1 in accordance with 1:5,000 RBI map accuracy standard. Therefore, this study concludes that geometrically, the building line resulted from the regularization is appropriate as feature in 1:5,000 RBI.