DETEKSI JENIS RETAK KULIT BUAYA PADA JALAN ASPAL MENGGUNAKAN CITRA DIGITAL

Abstract

          Alligator cracking, also known as fatigue cracking, is a series of interconnected cracks that occur on the surface of asphalt concrete layers as a result of fatigue failure caused by repeated vehicle loads. This type of cracking is classified into three levels, namely low, medium, and high severity. However, the detection of road cracking is still commonly performed manually, which becomes an obstacle in accurately identifying alligator cracking conditions. To address this issue, this study develops an algorithm using YOLOv11 implemented in an Android application to facilitate the identification of low, medium, and high alligator cracking. The proposed method includes dataset collection, data labeling, preprocessing, data augmentation, YOLOv11 model training, implementation of the TensorFlow Lite (tflite) model into the application, and real-time testing. The experimental results show that the developed model achieves a Precision of 83.4 percent, Recall of 83.3 percent, mAP50 of 83.1 percent, mAP50–90 of 59.3 percent, F1-Score of 83.3 percent, and an indirect testing accuracy above 90 percent. Based on these results, it can be concluded that the developed application is capable of accurately identifying alligator cracking in real time.

          Retak kulit buaya atau retak lelah merupakan rangkaian retak saling berhubungan pada permukaan lapis beton aspal sebagai akibat keruntuhan lelah oleh beban kendaraan yang berulang. Retak ini juga diklasifikasikan menjadi 3 bagian yaitu retak rendah, sedang, dan tinggI. Namun dalam deteksi jalan retak tersebut masih dilakukan secara manual, sehingga hal ini menjadi salah satu hambatan dalam mengidentifikasi jalan retak kulit buaya. Untuk menjawab pertanyaan tersepu peneliti membuat algoritma menggunakan YOLOV11 pada aplikasi android untuk memudahkan identifikasi jalan retak kulit buaya rendah, sedang, dan tinggi. Metode yang digunakan meliputi pengumpulan dataset, pelabelan, preprocessing, augmentasi, pelatihan model YOLOV11, pengaplikasian model.tflite kedalam apliaksi, juga pengujian secara realtime. Sehingga didapatkan hasil hasil Presicion 83.4%, Recall 83.3%, mAP50 83.1%, mAP50-90 59.3%, F1-Score 83.3%, dan akurasi pengujian tidak langsung >90%. kesimpulannya aplikasi mampu mengidenttifikasi secara realtime dengan hasil yang akurat.