Electronic Theses and Dissertation

Hama tikus (Rattus argentiventer) merupakan penyebab utama kehilangan hasil panen padi pascapanen di Indonesia, dengan tingkat kerugian yang dapat mencapai 37% per hektar serta penurunan kualitas gabah akibat kontaminasi. Metode pengendalian konvensional seperti rodentisida dan perangkap sering kali kurang efektif dalam skala besar dan berpotensi menimbulkan dampak negatif terhadap lingkungan. Penelitian ini bertujuan mengembangkan sistem deteksi tikus secara real-time yang inovatif, efisien, dan berkelanjutan menggunakan algoritma deep learning YOLOv11 yang diimplementasikan pada Raspberry Pi 5, untuk mendukung pemantauan dini di lingkungan penyimpanan padi. Sebanyak 410 citra tikus dikumpulkan dari Google Image, kemudian disegmentasi menggunakan Roboflow. Dataset dibagi menjadi 328 data latih, 41 validasi, dan 41 uji. Model YOLOv11 dilatih hingga 500 epoch dengan parameter early stopping 100, learning rate 0.002, batch size 4, dan optimizer AdamW. Model selanjutnya diintegrasikan pada perangkat Raspberry Pi 5 yang dilengkapi webcam dan buzzer. Hasil pelatihan menunjukkan model mencapai presisi 0.955, recall 0.86, mAP@0.50 sebesar 0.87, dan mAP@0.50–0.95 sebesar 0.39. Meskipun performa pada data latih tinggi, teridentifikasi overfitting pada data validasi mulai dari epoch ke-40 hingga ke-50. Prototipe berhasil mendeteksi tikus secara real-time dalam simulasi dan mengaktifkan buzzer sebagai peringatan. Kesimpulannya, sistem deteksi berbasis YOLOv11 pada Raspberry Pi 5 menunjukkan efektivitas untuk deteksi dini hama tikus, namun diperlukan optimasi lebih lanjut guna meningkatkan presisi lokalisasi dan mengurangi overfitting untuk implementasi jangka panjang.

The rice field rat (Rattus argentiventer) is a primary cause of post-harvest rice yield losses in Indonesia, with potential damage reaching up to 37% per hectare, as well as a decline in grain quality due to contamination. Conventional control methods such as rodenticides and traps are often ineffective on a large scale and can pose environmental risks. This study aims to develop an innovative, efficient, and sustainable real-time rat detection system using the YOLOv11 deep learning algorithm implemented on a Raspberry Pi 5, supporting early monitoring in paddy storage environments. A total of 410 rat images were collected from Google Images and segmented using Roboflow. The dataset was divided into 328 training images, 41 validation images, and 41 testing images. The YOLOv11 model was trained for up to 500 epochs with an early stopping threshold of 100, a learning rate of 0.002, a batch size of 4, and the AdamW optimizer. The trained model was then integrated into a Raspberry Pi 5 device equipped with a webcam and a buzzer. The training results showed that the model achieved a precision of 0.955, recall of 0.86, mAP@0.50 of 0.87, and mAP@0.50–0.95 of 0.39. Although the model performed well on the training set, signs of overfitting were observed in the validation set between epochs 40 and 50. The prototype successfully detected rats in real-time during simulations and activated the buzzer as an alert. In conclusion, the YOLOv11-based detection system implemented on Raspberry Pi 5 demonstrates effectiveness in early rat pest detection. However, further optimization is needed to enhance localization precision and reduce overfitting for long-term deployment in real-world agricultural storage settings.