# 海龜也有身分證，這次是鱗片不是晶片

- URL: https://justfly.idv.tw/%e6%b5%b7%e9%be%9c%e4%b9%9f%e6%9c%89%e8%ba%ab%e5%88%86%e8%ad%89%ef%bc%8c%e9%80%99%e6%ac%a1%e6%98%af%e9%b1%97%e7%89%87%e4%b8%8d%e6%98%af%e6%99%b6%e7%89%87/
- 日期: 2026-10-02
- 分類: 我知故我在
- 標籤: AI辨識, 公民科學, 台灣生態, 海洋保育

![海龜也有身分證，這次是鱗片不是晶片]
海龜的臉，眼睛周圍一圈鱗片,形狀、大小、排列方式,每一隻都長得不一樣。像指紋,但長在臉上。過去辨識這些紋路得靠人工一張張比對照片,現在這件事交給了一套叫「TurtleFinder」的系統。

9月16日,台灣海洋保育署聯合宏碁及海龜保育團體正式啟用這個平台。操作很直接:上傳一張海龜臉部照片,系統自動分析眼周與臉側鱗片特徵,拿去資料庫比對,確認是否為舊識。背後整合的是「TurtleSpot Taiwan」這個公民科學計畫累積多年的4,000筆紀錄,涵蓋綠蠵龜、玳瑁與欖蠵龜,辨識出的個體數近900隻。

##### 人工比對的極限

海龜保育一直有個笨重的環節。潛水員拍到一隻海龜,想知道牠是不是去年在同一片海域見過的那隻,得翻照片庫、憑肉眼找相似鱗片排列。資料量一旦累積到幾千筆,人力比對就是瓶頸。[機器學習](https://zh.wikipedia.org/wiki/%E6%A9%9F%E5%99%A8%E5%AD%B8%E7%BF%92)擅長的正是這種重複性高、模式化的辨識工作，把原本要花幾小時的比對,壓縮到幾秒鐘。

台灣並非第一次把人工智慧用在生態監測上。水下攝影機辨識魚類與珊瑚健康狀況、環境DNA技術分析海水樣本找出稀有物種分布,都是同一條路線上的嘗試。TurtleFinder的不同之處在於,辨識對象換成了一種「長相本身就是身分證」的動物——海龜的臉部鱗片終其一生幾乎不變,這讓AI辨識有了穩定的生物學基礎,而不只是影像比對的技術展示。

##### 資料庫比演算法更關鍵

一套辨識系統再聰明,沒有足夠的比對樣本也是空轉。TurtleFinder能跑起來,靠的是過去累積的4,000筆公民科學紀錄——也就是潛水員、海洋愛好者多年來自發回報的目擊紀錄。這批資料原本分散在不同單位手中,格式不一、難以互通。這次的整合,等於是把散落各處的觀察記錄,收攏成一個統一的海龜身分資料庫。

台灣在公民科學上並非新手。珊瑚礁體檢計畫從2009年開始,每年有數百名志工參與調查；鯨豚研究也長期仰賴賞鯨遊客與漁民回報目擊資訊建立資料庫。TurtleFinder延續的是同一套邏輯——保育機構缺的從來不是熱情參與者,而是把分散觀察整理成可用資料的工具。

##### 按下快門,就是做研究

這個平台真正改變的,是「參與」這件事的門檻。過去,一般人頂多能做的是通報目擊地點,辨識個體身份是專業人員的工作。現在系統開放公眾上傳照片、回報目擊紀錄,一次潛水拍下的一張臉部特寫,就可能補進資料庫、幫忙確認某隻海龜是否再度出現在同一片海域。

海龜是會跨國界洄游的動物,牠們不會管哪片海域屬於誰的保護區。資料庫累積得越完整,能追蹤的個體生命史就越長——從某次目擊的幼龜,到幾年後重新出現的成龜,中間那段看不見的日子,正是靠這類零散回報串起來的。4,000筆紀錄、近900隻個體,這是起點,資料庫會隨著每一次上傳繼續長大。

— 鄭佩玲

### Sea Turtles Now Have an ID Card, Made of Scales Not Chips

A sea turtle’s face carries a ring of scales around each eye. Their shape, size, and arrangement differ from one turtle to the next — like a fingerprint, except it sits on the face. Matching these patterns used to mean scrolling through photo archives by hand. Now a platform called TurtleFinder does the matching instead.

Taiwan’s Ocean Conservation Administration, working with Acer and sea turtle conservation groups, launched the platform on September 16. The process is simple: upload a photo of a turtle’s face, and the system analyzes the shape and arrangement of scales around the eyes and cheeks, then checks it against a database to see if the turtle has been recorded before. The database draws on four years’ worth of citizen science — a project called TurtleSpot Taiwan — which had already accumulated 4,000 records covering green sea turtles, hawksbills, and olive ridleys, identifying nearly 900 individual turtles.

##### Where Manual Matching Breaks Down

Sea turtle conservation has always had one clunky step. A diver photographs a turtle and wants to know whether it’s the same one spotted in the same waters last year. That meant digging through photo libraries and comparing scale patterns by eye. Once records pile up into the thousands, human matching becomes the bottleneck. [Machine learning](https://en.wikipedia.org/wiki/Machine_learning) is suited exactly to this kind of repetitive, pattern-based task — compressing hours of comparison into seconds.

This isn’t Taiwan’s first attempt at applying AI to ecological monitoring. Underwater cameras already help identify fish species and assess coral reef health; environmental DNA analysis of seawater samples helps locate rare species. What sets TurtleFinder apart is the subject itself — a turtle’s facial scales stay largely unchanged for life, giving the AI a stable biological basis to work from rather than just a clever image-matching trick.

##### The Database Matters More Than the Algorithm

No matter how sharp the recognition model is, it’s useless without enough reference data. TurtleFinder runs on four years of citizen-collected sightings — divers and ocean enthusiasts voluntarily reporting what they’d seen. That data used to sit scattered across different organizations, in incompatible formats. This launch consolidates it into one unified turtle identity database.

Taiwan isn’t new to citizen science. A coral reef survey program has run annually since 2009 with hundreds of volunteers; whale and dolphin researchers have long relied on sighting reports from whale-watching tourists and fishermen to build their databases. TurtleFinder follows the same logic — conservation agencies were never short on willing participants, just short on tools to turn scattered observations into usable data.

##### A Photo Becomes a Data Point

What the platform really changes is the threshold for participation. Previously, an ordinary person’s contribution stopped at reporting a sighting location; identifying the individual turtle was work for trained specialists. Now the public can upload photos and report sightings directly — one close-up shot from a single dive might confirm that a particular turtle has returned to the same stretch of water.

Sea turtles migrate across borders; they don’t recognize whose protected area they’re swimming through. The more complete the database grows, the longer an individual turtle’s life history can be tracked — from a hatchling spotted once to an adult resurfacing years later in the same reef. The gap in between gets filled in by exactly this kind of scattered, crowd-sourced reporting. 4,000 records and nearly 900 individuals mark a starting point — the database grows with every new upload.

— 鄭佩玲