Electronic Theses and Dissertation

Penelitian ini mengembangkan sistem deteksi hama kecoa otomatis pada rumah burung walet dengan mengintegrasikan algoritma YOLOv10, RT-DETR, dan SSD-VGG-16, serta teknologi Internet of Things (IoT) dan Telegram Bot untuk notifikasi. Sistem ini bertujuan meningkatkan akurasi deteksi, mengurangi inspeksi manual, dan meminimalkan gangguan terhadap burung walet. Dataset dibangun dari rekaman CCTV inframerah dan diperluas melalui augmentasi citra. Evaluasi dilakukan terhadap tujuh varian YOLOv10, serta model pembanding YOLOv8-m, RT-DETR-l, dan SSD-VGG-16 menggunakan metrik precision, recall, F1-score, dan mAP50. Hasil menunjukkan YOLOv10-s memiliki F1 Score tertinggi (0,9851), YOLOv10-x mencatat precision tertinggi (0,9900), dan YOLOv10-b unggul dalam recall (0,9869). RT-DETR-l juga menunjukkan performa tinggi dengan recall 0,9824 dan F1 Score 0,9846, sedangkan SSD-VGG-16 memiliki precision tinggi meskipun jumlah deteksinya paling sedikit. Seluruh model berhasil mengirimkan hasil deteksi ke Telegram Bot tanpa kehilangan notifikasi. Sistem dilengkapi fitur Region of Interest (ROI) dan hanya mengirim notifikasi saat jumlah hama meningkat, dengan reset otomatis setiap 24 jam. Hasil ini menunjukkan bahwa sistem yang dikembangkan efektif dan andal untuk mendeteksi hama kecoa secara responsif di rumah burung walet

This study develops an automated pest detection system for swiftlet houses using YOLOv10, RT-DETR, and SSD-VGG-16 algorithms, integrated with IoT technology and Telegram Bot notifications. The main objectives are to improve detection accuracy, reduce manual inspections, and minimize disturbance to swiftlet activity. A custom dataset was created from infrared CCTV footage and enhanced through image augmentation. The system was evaluated using seven YOLOv10 variants, YOLOv8-m, RT-DETR-l, and SSD-VGG-16, based on precision, recall, F1-score, and mAP50. YOLOv10-s achieved the highest F1 Score (0.9851), YOLOv10-x had the highest precision (0.9900), and YOLOv10-b recorded the highest recall (0.9869), showing overall superiority among tested models. RT-DETR-l also showed competitive results, with an F1 Score of 0.9846 and recall of 0.9824, despite its larger model size. The ROI (Region of Interest) feature effectively improved accuracy by focusing detection on key areas, reducing false positives. The system’s Telegram Bot successfully sent all detection alerts with 100% message delivery across all models. Notifications were triggered only when pest counts increased, supported by a reset mechanism to prevent redundancy. The results confirm the system’s reliability and efficiency, making it a promising solution for real-time pest monitoring in swiftlet farming environments