| 摘要: | 檳榔產業曾經是台灣產值排名第二的農業經濟作物，產值僅次於稻米。由於檳榔主要種植於山坡地，因此被政府視為土石流的原因之一，加上其具有致癌性，因此政府開始推行多種相關政策來縮減種植面積。而在政策實施期間，因涉及檳榔樹的廢園補助，需要盤點檳榔樹的數量，然而目前傳統的計算方式仍然是由人工逐棵計算，無法在短時間內完成計算，因此本研究提出利用無人機收集高解析度影像，結合深度學習中的物件偵測技術，並比較比較了one stage、two stage與anchor-free三種架構中較具代表性的物件偵測方法 (Efficientdet、Faster R-CNN與CenterNet) ，以自動化的方式進行檳榔樹的辨識與計數，取代傳統人工計算方法。在物件偵測方法部分，結果顯示，在物件偵測方法部分，本研究比較了Center Net、Efficientdet與Faster R-CNN三種不同架構的方法，其中表現最佳的方法為anchor-free架構中的Center Net，經過測試，Center Net的Recal與Accuracy可達到82.8%，Precision的部分則可達到100%，在面對不同背景、不同種植密度、不同亮度、不同地面解析度的影像時，Center Net都可以成功辨識出大部分的檳榔樹，在山區檳榔樹辨識方面有良好的應用性。
The betel nut industry used to be Taiwan's second-largest agricultural economic crop in terms of production value, second only to rice. However, due to betel nut cultivation mainly occurring on sloped lands, it was considered one of the factors contributing to soil erosion and landslides. Additionally, betel nuts have been linked to carcinogenic properties. As a result, the government initiated various policies to reduce the cultivation area of betel nuts.
During the implementation of these policies, as part of the reimbursement for converting betel nut gardens into other uses, it became necessary to inventory the number of betel nut trees. However, the traditional manual counting method proved time-consuming, as it required counting each tree individually. Therefore, this study proposed an automated approach using high-resolution imagery collected by unmanned aerial vehicles (UAVs). This approach combined object detection techniques from deep learning to recognize and count betel nut trees efficiently. Three representative object detection methods, namely Efficientdet, Faster R-CNN, and CenterNet, which belong to one stage, two stage, and anchor-free architectures, were compared in this research.
The results indicated that the anchor-free CenterNet approach performed the best among the three methods. Through testing, CenterNet achieved a Recall and Accuracy of 82.8% and a Precision of 100%. It successfully identified most of the betel nut trees even in diverse image backgrounds, different planting densities, various lighting conditions, and varying ground resolutions. The application of CenterNet in betel nut tree recognition in mountainous regions proved to be highly effective. |
