Electronic Theses and Dissertation

Ekosistem mangrove memiliki peran penting dalam mitigasi perubahan iklim karena kemampuannya dalam menyerap dan menyimpan karbon dioksida (co₂). estimasi daya serap karbon mangrove umumnya masih dilakukan secara manual, yang memerlukan waktu dan sumber daya yang besar. penelitian ini bertujuan untuk mengembangkan metode estimasi daya serap karbon mangrove secara otomatis menggunakan model deep learning mobilenetv2 berbasis deteksi objek. data yang digunakan berupa citra udara kawasan mangrove gampong pande, banda aceh, yang terdiri atas tiga genus utama, yaitu rhizophora, avicennia, dan nypa. citra dianotasi secara manual dengan bounding box dan digunakan untuk melatih model deteksi objek. evaluasi performa model dilakukan dengan mengukur nilai average precision (ap), average recall (ar), dan intersection over union (iou). hasil evaluasi menunjukkan bahwa performa deteksi model masih rendah, terutama pada objek dengan ukuran kecil dan tumpang tindih, dengan nilai ap dan ar yang belum mencapai standar akurasi yang diharapkan. estimasi daya serap karbon berdasarkan hasil deteksi model menunjukkan deviasi yang cukup besar jika dibandingkan dengan estimasi dari ground truth maupun pengukuran lapangan. validasi menggunakan tiga pendekatan: prediksi model, ground truth, dan pengukuran lapangan menunjukkan bahwa anotasi manual dan data lapangan memiliki kesesuaian tinggi, sementara hasil model mengalami selisih estimasi sebesar 30–45% pada beberapa objek, khususnya avicennia marina dan nypa fruticans. hal ini mengindikasikan bahwa kelemahan utama terletak pada akurasi spasial model. penelitian ini merekomendasikan peningkatan kualitas data, eksplorasi arsitektur model lanjutan, serta integrasi dengan data lapangan untuk pengembangan lebih lanjut.

Mangrove ecosystems play a vital role in climate change mitigation due to their ability to absorb and store carbon dioxide (CO₂). However, the estimation of mangrove carbon sequestration is generally still conducted manually, which is time-consuming and resource-intensive. This study aims to develop an automatic estimation method for mangrove carbon sequestration using the MobileNetV2 deep learning model based on object detection. The data used consisted of aerial imagery of mangrove areas in Gampong Pande, Banda Aceh, which included three main genera: Rhizophora, Avicennia, and Nypa. The images were manually annotated using bounding boxes and used to train the object detection model. Model performance was evaluated using Average Precision (AP), Average Recall (AR), and Intersection over Union (IoU) metrics. The evaluation results show that the model’s detection performance remains low, especially for small and overlapping objects, with AP and AR values falling below expected accuracy standards. Carbon sequestration estimates derived from the model's detection results exhibited significant deviation when compared to those from ground truth annotations and field measurements. Validation using three approaches: model predictions, ground truth annotations, and field-based calculations indicated that ground truth and field data had a high degree of agreement, while the model's estimations showed deviations of 30–45% for several objects, particularly Avicennia marina and Nypa fruticans. These findings suggest that the primary limitation lies in the spatial detection accuracy of the model. This study recommends improving data quality, exploring more advanced model architectures, and integrating field data for further development.