Deep Object Detection Approaches for Identification of Seagrass Species Using Faster R-CNN and Single Shot MultiBox Detector (SSD)

Abstrak

Abstract Seagrass meadows play vital ecological roles by supporting biodiversity, stabilizing sediments, and maintaining coastal productivity. However, species-level monitoring in tropical waters remains difficult due to high turbidity, uneven illumination, and overlapping morphology that hinder visual identification. Traditional field surveys are time-consuming and inconsistent, underscoring the need for automated and scalable approaches. Deep learning–based object detection offers promising solutions but remains underexplored in tropical seagrass ecosystems. This study employs Faster R-CNN and Single Shot MultiBox Detector (SSD) on a unique in-situ underwater dataset from the Riau Islands, Indonesia, to establish baseline detection performance under real environmental conditions. A total of 1,480 images were enhanced using CLAHE, LAB color correction, and sharpening, then divided into training and validation sets. Model performance was assessed using mAP at IoU thresholds 0.50 and 0.75, precision, recall, and inference time. Faster R-CNN achieved higher mAP (38.69%) than SSD (31.35%) but required longer processing. These findings highlight trade-offs between accuracy and speed and form the first baseline for automated seagrass detection in Indonesia.

Identifier

• OpenAlex: https://openalex.org/W4417165348
• DOI: https://doi.org/10.1088/1755-1315/1559/1/012012