Electronic Theses and Dissertation

Radio frequency interference (rfi) pada sistem radar cuaca kerap menimbulkan distorsi pada citra, sehingga dapat menurunkan akurasi analisis meteorologi. hingga saat ini, proses identifikasi gangguan masih dilakukan secara manual oleh operator, yang menimbulkan tantangan dari sisi konsistensi, kecepatan, dan skalabilitas. penelitian ini berfokus pada implementasi beberapa varian you only look once versi 12 (yolov12) serta perbandingannya dengan model real-time detection transformer (rt-detr) dari pustaka ultralytics untuk mendeteksi rfi secara otomatis. berbeda dengan sebagian besar penelitian sebelumnya yang menggunakan dataset radar cuaca yang berasal dari luar negeri, penelitian ini membangun dataset representatif berbasis citra radar cuaca indonesia, yang disusun dari contoh citra terganggu dan citra bersih agar lebih sesuai dengan karakteristik spektral serta kondisi operasional radar nasional. model deep learning dilatih menggunakan dataset tersebut untuk meningkatkan akurasi deteksi sekaligus kemampuan adaptasi terhadap lingkungan radar indonesia. evaluasi kinerja dilakukan menggunakan metrik precision, recall, mean average precision (map), f1-score, dan efisiensi komputasi guna mengidentifikasi pendekatan paling efektif untuk tugas deteksi rfi. hasil penelitian menunjukkan bahwa seluruh varian yolov12 memiliki performa lebih unggul dibandingkan rt-detr dalam mendeteksi rfi pada citra radar cuaca. varian yolov12x memperoleh akurasi tertinggi dengan nilai map@0.5 sebesar 0,8269, sementara yolov12m dan yolov12l menawarkan keseimbangan yang baik antara akurasi dan efisiensi komputasi. temuan ini mengindikasikan bahwa yolov12 merupakan solusi yang lebih andal dan efisien untuk kebutuhan pemantauan operasional radar cuaca. kata kuci: radio frequency inteference, citra radar cuaca, deep learning, yolov12, rt-detr

Abstract—Radio frequency interference (RFI) in weather radar systems frequently leads to image distortions, undermining the accuracy of meteorological analyses. Currently, interference detection relies on manual inspection by human operators, which poses challenges in terms of consistency, speed, and scalability. This study focuses on the implementation of multiple You Only Look Once version 12 (YOLOv12) variants and compares their performance with the real-time detection transformer (RT-DETR) model from the Ultralytics library for automatic detection tasks. While most previous research has employed foreign weather radar datasets, these may not be compatible with the spectral characteristics and operational conditions of radar systems in Indonesia. To address this gap, the study focuses on developing and leveraging a representative weather radar image dataset tailored to the Indonesian context, including examples of RFI and clean radar imagery. The deep learning models were trained on this dataset to enhance detection accuracy and improve adaptation to local radar environments. Model performances were evaluated using precision, recall, mean average precision (mAP), F1-score, and computational efficiency to identify the most effective approach for detecting RFI. Results show that the YOLOv12 models outperform RT-DETR in detecting RFI on weather radar imagery. The YOLOv12x variant achieved the highest accuracy of 0.828 (mAP@0.5), while YOLOv12m and YOLOv12l offered a balanced trade-off between accuracy and efficiency, indicating that YOLOv12 provides a more reliable and computationally efficient solution for operational weather radar monitoring. Keywords—radio frequency interference, weather radar imagery, deep learning, YOLOv12, RT-DETR