Claim Missing Document
Check
Articles

Found 1 Documents
Search
Journal : https://journal.amikveteran.ac.id/index.php/teknik

DETEKSI BANJIR AREA PERKOTAAN BERBASIS CITRA DIGITAL CONVOLUTIONAL NEURAL NETWORK (VGG19) Habibullah Akbar; Diah Aryani; Muhamad Bahrul Ulum
Jurnal Teknik Mesin, Elektro dan Ilmu Komputer Vol. 2 No. 3 (2022): November : Jurnal Teknik Mesin, Elektro dan Ilmu Komputer
Publisher : Amik Veteran Porwokerto

Show Abstract | Download Original | Original Source | Check in Google Scholar | Full PDF (992.245 KB) | DOI: 10.55606/teknik.v2i3.798

Abstract

Geographically and demographically, Indonesia has natural conditions that have the potential for floods disaster. There are at least 16,771 islands and 65,017 rivers that fill the archipelago. Unfortunately, the ever-increasing urban population accompanied by a lack of awareness and preparation for protecting the environment has resulted in a higher risk of flooding in urban areas. This study utilizes digital imagery to detect flood conditions in urban areas. In terms of access, digital images are available in urban CCTV monitoring systems as well as office areas, housing, and from people who have smartphones. The detection method used in this study is the VGG19 model which consists of 16 convolution layers and 3 standard classification layers. All convolution layers are divided into 5 blocks followed by a MaxPooling layer for each block to reduce the resolution of the input image. In the last layer, SoftMax layer is used to estimate the probability between flood labels and normal conditions. There are 4 parameters that were optimized during the VGG19 model training process, namely Batch Size, Learning Rate, Dropout and Epoch (training repetition). To test the proposed model, public datasets are used, namely the Roadway Flooding Image Dataset and Road Vehicle Images Dataset. The best flood detection results (or normal conditions) achieve the accuracy of 98.78%. As for the other three performance metrics, namely precision, recall and F1-score, they reach 99%. These results are generated by the VGG19 model with a Batch Size parameter of 20, a Learning Rate of 1e-5 (0.00001), 50% Dropout and 100 Epoch. The achievement values of the four metrics can be considered quite good, so that the VGG19 model has the opportunity to be developed for flood detection applications in order to monitor urban flood conditions.