Detail Cantuman

Kebakaran hutan merupakan ancaman serius yang merusak ekosistem dan meningkatkan emisi karbon monoksida hingga 30% di atmosfer. Karbon monoksida bersifat sangat beracun karena memiliki afinitas lebih kuat terhadap hemoglobin dibandingkan oksigen, sehingga menghambat distribusi oksigen dan berpotensi menimbulkan hipoksia jaringan. Untuk memungkinkan respons cepat di wilayah rawan kebakaran, penelitian ini mengembangkan sistem deteksi potensi kebakaran hutan berbasis citra UAV menggunakan model YOLOv10. Indikator potensi kebakaran ditetapkan apabila objek api dan asap terdeteksi secara bersamaan. Dataset yang digunakan berasal dari gabungan Roboflow Universe dan Kaggle dengan dua kelas objek, yaitu fire dan smoke. Model dilatih menggunakan enam jenis optimizer, dua variasi batch size, dan dua nilai learning rate, menghasilkan 24 konfigurasi pelatihan. Model terbaik diperoleh dengan optimizer NAdam, batch size 24, dan learning rate 0.001 dengan nilai accuracy 0.879, precision 0.8705, recall 0.8575, F1 score 0.863, mAP 0.870, dan mean IoU 0.7373. Pengujian real-time menggunakan drone DJI Mini 4 Pro menunjukkan sistem berjalan stabil, responsif, dan mampu menampilkan peringatan “berpotensi kebakaran” secara akurat ketika api dan asap terdeteksi bersamaan. Hasil ini mengonfirmasi efektivitas YOLOv10 dalam mendeteksi potensi kebakaran hutan secara real-time, terutama dengan dukungan optimasi hyperparameter yang tepat.

Forest fires are a serious threat that damages ecosystems and increases carbon monoxide emissions by up to 30% in the atmosphere. Carbon monoxide is highly toxic because it has a stronger affinity for hemoglobin than oxygen, thus inhibiting oxygen distribution and potentially causing tissue hypoxia. To enable rapid response in fire-prone areas, this study developed a UAV imagery-based forest fire potential detection system using the YOLOv10 model. A potential fire indicator is determined when fire and smoke objects are detected simultaneously. The dataset used comes from a combination of Roboflow Universe and Kaggle with two object classes, namely fire and smoke. The model was trained using six types of optimizers, two variations of batch sizes, and two learning rate values, resulting in 24 training configurations. The best model was obtained with the NAdam optimizer, batch size 24, and a learning rate of 0.001 with an accuracy of 0.879, precision 0.8705, recall 0.8575, F1 score 0.863, mAP 0.870, and mean IoU 0.7373. Real-time testing using a DJI Mini 4 Pro drone demonstrated the system's stability, responsiveness, and ability to accurately display a "potential fire" warning when fire and smoke are detected simultaneously. These results confirm YOLOv10's effectiveness in detecting potential forest fires in real time, especially with the support of appropriate hyperparameter optimization.