UV-NDVI與雙目立體視覺技術在環控農牧業智慧監測中的開發與應用

Abstract

本研究旨在開發並應用視覺技術與機制模型於環控農牧業的智慧監測與品質優化。研究涵蓋三個相互關聯的主題： 首先，研究探討利用 UV-A 光照作為逆境因子，以提升紅萵苣的營養價值與商業品質。透過兩階段篩選至最佳化框架，實驗結果顯示適當的 UV-A 照射策略（每日 6 小時，持續 6 天，稱為 L6D6 配方）可在不影響鮮重的前提下，有效促進 'Lollo Rosso' 萵苣轉色並提高花青素含量 14%、每株花青素總量提升 19.9%。為深入理解並預測 UV-A 照射對萵苣生長與品質的交互影響，本研究進一步開發了一個六狀態常微分方程（ODE）機制模型，整合了活性氧（ROS）動態、逆境損傷-修復過程，以及碳資源在生長與抗氧化防禦之間的競爭機制。模型在獨立驗證集上達到高準確度（12 項指標中 11 項誤差＜10%），並通過全域最佳化預測出理論最佳配方：每日 9 小時照射持續 4 天，可使花青素總量提高約 38.3%，同時對鮮重影響極小（-2.4%）。 其次，基於 UV-A 與植物螢光的交互作用，本研究開發了 UV-NDVI（UV induced normalized difference vegetation index，紫外光誘導差異植生指標）影像監測技術。實驗證明，UV-NDVI 能比傳統 SI-NDVI 及肉眼觀察更早、更靈敏地偵測到由水分逆境引起的植物健康變化。

最後，作為環控農業監測的延伸應用，本研究建立了一套雙目立體視覺系統，結合 YOLOv4 與橢圓擬合演算法，用於非接觸式預測肉雞體重。在商業養殖場的驗證中，系統展現了高穩定性，經統計學校正後，平均相對誤差僅為2.28%。 本論文成功整合了機制建模、光譜影像監測與立體視覺等技術，為植物工廠的品質調控、作物健康預警及智慧畜牧管理提供了創新的解決方案。

This study develops and applies vision techniques and mechanistic modeling for intelligent monitoring and quality optimization in controlled environment agriculture (CEA) and animal husbandry. The research encompasses three interrelated topics: First, UV-A irradiation was investigated as a stress factor to enhance the nutritional value and commercial quality of red lettuce. Through a two-stage screening-to-optimization framework, results demonstrate that an appropriate UV-A strategy (6 h/day for 6 days; L6D6 treatment) effectively promotes coloration and increases anthocyanin concentration by 14% and total anthocyanin per plant by 19.9% without compromising fresh weight in 'Lollo Rosso' lettuce. To understand and predict the interactive effects of UV-A on lettuce growth and quality, a six-state ordinary differential equation (ODE) mechanistic model was developed, integrating reactive oxygen species (ROS) dynamics, stress damage-repair processes, and carbon resource competition between growth and antioxidant defense (Growth-Differentiation Balance Hypothesis). The model achieved high accuracy on an independent validation set (11 of 12 metrics < 10% error), and global optimization predicted a theoretical optimal regimen of 9 h/day for 4 days, projected to increase total anthocyanin by approximately 38.3% with minimal fresh weight impact (-2.4%). Second, building on the interaction between UV-A and plant fluorescence, a UV-NDVI (UV induced normalized difference vegetation index) imaging technique was developed. Experiments demonstrated that UV-NDVI detects water stress-induced changes in plant health earlier and more sensitively than traditional SI-NDVI and visual observation. Finally, as an extended application in CEA monitoring, a binocular stereo vision system combined with YOLOv4 and elliptic fitting algorithms was established for non-contact broiler chicken weight prediction. Commercial farm validation demonstrated high stability with a mean relative error of only 2.28% after statistical correction. This dissertation successfully integrates mechanistic modeling, spectral imaging, and stereo vision technologies, providing innovative solutions for quality control in plant factories, crop health early warning, and smart livestock management.