運用深度學習語意分割技術量化臺灣亞熱帶森林細根影像

Abstract

全球氣候變化對熱帶森林具有特別深遠的影響，導致碳匯估算出現顯著變化。細根是這一循環當中迅速消長的指標，其容易受到即將到來不穩定擾動的影響，因此對其進行量化作業至關重要。以往的研究參與了植物根系的量化工作，但幾乎都是在受控制的環境當下進行的。同樣地，迅速的根系分析也很難建立，因為根系隱藏在不透明的土壤中，難以直接觀察和進行長期監測。目前，微根管技術在成本、安裝操作便捷性和收集野外細根動態變化的能力之間取得了平衡。 本研究強調了我們改進的U-Net—FRS-Net，在微根管圖像中細根分割的創新性。實驗樣區選擇以台灣的棲蘭、福山及蓮華池三個亞熱帶森林地區做為代表。我們的修改包括額外的遺忘層塊應用和影像的預處理步驟，旨在提高根系分割任務中的精度和F1分數。我們將FRS-Net與兩種前人研究已提出的方法進行比較：SegRoot，一種具有可控網絡深度和寬度的深度學習架構，以及另一種工具—saRIA，一種採用自適應閾值和形態處理的方法。結果表明，FRS-Net在F1及IoU分數方面，在些許特定的資料集中，皆較優於其他方法。此外，我們的方法在整體平均精度方面達到了最高，有效地減少了假陽性，同時保持了高的真陽性率。

Global climate change has especially profound effects on tropical forests, leading to significant variations in carbon sink estimation. Fine roots, a rapid turnover index in this cycle, are easily affected by upcoming unstable disturbances, and it’s crucial to create quantification with them. Previous research has taken part in quantifying plant roots while almost all were deployed in stable environments. Similarly, rapid root system analysis is complex to build because the root system is hidden in non-transparent soil, which makes it challenging to observe directly in long-term monitoring. Currently, the minirhizotrons method balances cost, ease of deployment, and the ability to collect the dynamic variations of the natural fine root. This study underscores the novelty of our modified U-Net, FRS-Net, in fine root segmentation from minirhizotrons images. For the selection of experimental sites, we chose regions of Chi-Lan, Fu-Shan, and Lien-Hua-Chih as representatives of Taiwan‘s subtropical forests. Our modifications, including additional layers and specialized preprocessing steps, are designed to enhance the precision and F1 score in root segmentation tasks. We compared FRS-Net with two established methods: SegRoot, a deep learning-based architecture with controllable network depth and width, and saRIA, which employs adaptive thresholding and morphological operations. The results demonstrate that FRS-Net surpasses other methods in terms of F1 & IoU score in some certain datasets. Moreover, our method achieves the highest overall precision, effectively reducing false positives while maintaining high true positive rates. The integration of FRS-Net, along with SegRoot and saRIA, has made further advancements in the segmentation of fine roots in subtropical forests. These combined techniques will facilitate more detailed and reliable research in ecological and botanical studies. The findings suggest that leveraging advanced neural network architectures with tailored modifications can significantly improve the analysis of fine root systems. Additionally, the framework for segmenting fine roots in subtropical forests offers new perspectives on carbon sink quantification.