基於深度學習的珊瑚正射影像辨識

Abstract

珊瑚覆蓋率對於維持海洋生物多樣性與氣候調節具有關鍵作用，石珊瑚與軟珊瑚的分布變化亦常反映海洋污染與環境惡化的狀況，因此同時監測兩者對評估珊瑚礁健康尤為重要。本研究針對傳統穿越線法(LIT)在監測範圍、主觀性及資料保存上的限制，並考量過往研究多聚焦於石珊瑚的不足，提出一套結合水下錄影、攝影測量與語意分割模型的整合性監測方法。透過高解析度正射影像建置與自動化分類流程，本研究能有效辨識石珊瑚與軟珊瑚的空間分布。實驗結果顯示， 僅需 18 分鐘錄影時間，即可重建約 80 平方公尺樣區；語意分割模型訓練中，石珊瑚與軟珊瑚的 IoU 均達 97% 以上；測試影像中石珊瑚與軟珊瑚覆蓋率誤差分別為 -1.06% 與 -0.77%，符合常見珊瑚監測計畫中 ±5% 的誤差容忍度。本研究成功建立一套低成本、高效率且具客觀性的珊瑚監測架構。

Coral cover plays a vital role in maintaining marine biodiversity and regulating climate. Changes in the distribution of hard and soft corals often indicate marine pollution and environmental degradation, making the simultaneous monitoring of both essential for assessing reef health. This study addresses the limitations of traditional line-intercept transects (LIT), including restricted coverage, subjectivity, and lack of data preservation, by proposing an integrated method combining underwater video, photogrammetry, and semantic segmentation. Using high-resolution orthomosaics and automated classification, our approach effectively identifies the spatial distribution of both coral types. Field experiments show that an 18-minute video can reconstruct an 80 m2 area, with the trained model achieving IoUs over 97% for both hard and soft corals. The estimated cover errors in test images were −1.06% and −0.77%, well within the ±5% tolerance commonly accepted in coral monitoring. This study establishes a low-cost, efficient, and objective coral monitoring framework.