Electronic Theses and Dissertation

Abstrak- Mangrove berperan penting dalam melindungi pesisir, menyerap karbon, dan menjadi habitat berbagai spesies. Indonesia memiliki ekosistem mangrove terbesar di dunia, salah satunya di Gampong Pande, Banda Aceh. Namun, identifikasi jenis mangrove di lokasi ini masih dilakukan secara manual, sehingga memakan waktu, biaya besar, dan kurang efisien. Oleh karena itu, penelitian ini bertujuan untuk membandingkan kinerja model deteksi mangrove berbasis YOLOv8 dan Cascade R-CNN guna mengevaluasi efektivitas dan efisiensi masing-masing model dalam mendeteksi vegetasi mangrove pada citra udara. Studi ini dilakukan di Gampong Pande, Kecamatan Kutaraja, Banda Aceh, yang memiliki ekosistem mangrove dengan tiga genus utama: Rhizophora, Avicennia, dan Nypa fruticans. Data penelitian berupa citra drone resolusi tinggi dari vegetasi mangrove di lokasi tersebut. Evaluasi kinerja model dilakukan menggunakan berbagai metrik seperti akurasi, presisi, recall, dan mean Average Precision (mAP). Penelitian ini bertujuan untuk membandingkan dua model deteksi objek berbasis deep learning, yaitu YOLOv8 dan Cascade R-CNN, dalam mendeteksi vegetasi mangrove dari citra udara. Kedua model memiliki pendekatan arsitektur yang berbeda, di mana YOLOv8 dikenal sebagai model deteksi satu tahap (one-stage detector) yang mengutamakan kecepatan, sedangkan Cascade R-CNN merupakan model dua tahap (two-stage detector) yang menekankan pada presisi. Evaluasi dilakukan terhadap performa kedua model dalam mendeteksi berbagai jenis mangrove pada citra dengan kondisi lingkungan yang bervariasi, seperti perbedaan tekstur tanah, perubahan musiman, serta gangguan visual lainnya. Hasil evaluasi menunjukkan bahwa YOLOv8m unggul dari segi kecepatan dan efisiensi pelatihan dengan skor mAP50 sebesar 0,200. Namun, Cascade R-CNN meskipun mencatat skor mAP50 lebih rendah (0,168), mampu menghasilkan deteksi lebih stabil dan akurat dengan confidence score sebagian besar di atas 90%, serta konsisten mengenali label seperti Rhizophora tree dan Rhizophora sapling. Sebaliknya, YOLOv8m menghasilkan jumlah deteksi lebih sedikit dengan confidence score cenderung rendah di bawah 50%. Temuan ini menegaskan bahwa arsitektur two-stage seperti Cascade R-CNN lebih unggul dalam akurasi, sedangkan YOLOv8m lebih efisien. Dengan demikian, pemilihan model sebaiknya disesuaikan dengan kebutuhan, apakah mengutamakan efisiensi atau presisi, guna mendukung konservasi, pemantauan, dan rehabilitasi ekosistem mangrove secara efektif dan berbasis teknologi.

Kata Kunci: Mangrove, YOLOv8, Cascade R-CNN, Deep Learning, Deteksi Objek.

Abstract- Mangroves play an important role in protecting coastlines, absorbing carbon, and serving as habitats for various species. Indonesia has the largest mangrove ecosystem in the world, one of which is located in Gampong Pande, Banda Aceh. However, the identification of mangrove species in this area is still carried out manually, which requires significant time, high costs, and is less efficient. Therefore, this study aims to compare the performance of mangrove detection models based on YOLOv8 and Cascade R-CNN to evaluate the effectiveness and efficiency of each model in detecting mangrove vegetation from aerial imagery. This study was conducted in Gampong Pande, Kutaraja District, Banda Aceh, which hosts a mangrove ecosystem with three main genera: Rhizophora, Avicennia, and Nypa fruticans. The research data consisted of high-resolution drone imagery of mangrove vegetation in the area. The performance of the models was evaluated using various metrics such as accuracy, precision, recall, and mean Average Precision (mAP). The study aimed to compare two deep learning–based object detection models, YOLOv8 and Cascade R-CNN, in detecting mangrove vegetation from aerial imagery. The two models adopt different architectural approaches, where YOLOv8 is known as a one-stage detector that prioritizes speed, while Cascade R-CNN is a two-stage detector that emphasizes precision. Evaluation was carried out on the performance of both models in detecting various types of mangroves under diverse environmental conditions, such as soil texture variations, seasonal changes, and other visual disturbances. The evaluation results show that YOLOv8m outperformed in terms of speed and training efficiency with an mAP50 score of 0.200. However, Cascade R-CNN, despite recording a lower mAP50 score (0.168), was able to produce more stable and accurate detections, with most confidence scores above 90%, and consistently recognized labels such as Rhizophora tree and Rhizophora sapling. In contrast, YOLOv8m yielded fewer detections, with confidence scores tending to fall below 50%. These findings highlight that two-stage architectures such as Cascade R-CNN are superior in accuracy, whereas YOLOv8m excels in efficiency. Thus, the choice of model should be adjusted according to specific needs—whether prioritizing efficiency or precision—in order to support effective, technology-based mangrove conservation, monitoring, and rehabilitation. Keywords: Mangrove, YOLOv8, Cascade R-CNN, Deep Learning, Object Detection.