Sri Lankan seagrass families identification using underwater images through image enhancement and deep learning techniques

Abstract

Seagrass ecosystems play a vital role in marine biodiversity, carbon sequestration, and coastal protection, yet their monitoring is limited by challenges in underwater imaging and manual species identification. Automated classification of seagrass families is essential for large-scale, consistent, and timely ecological assessments. This study proposes a deep learning–based ensemble framework for classifying three predominant seagrass families in Sri Lanka’s coastal waters: Hydrocharitaceae, Cymodoceaceae, and Ruppiaceae. Over 700 raw underwater images were collected and processed using a two-stage enhancement pipeline to improve visibility and color fidelity for reliable learning. Multiple convolutional neural networks were evaluated, and a stacking ensemble integrating DenseNet121, VGG19, and MobileNetV2 achieved a 99% classification accuracy. The proposed framework advances automated seagrass identification and contributes to improved computer vision techniques for scalable marine monitoring and conservation.