應用自動化與人工智慧技術於瓢蟲規模化養殖 - 以小十三星瓢蟲幼蟲之影像辨識與餵食系統為例

Abstract

儘管農藥的使用為糧食生產帶來了顯著效益，但近年來其衍生之環境問題與健康問題，逐漸引發社會與學界的關注。因此，發展環境友善且具永續性的病蟲害管理策略已成為當前農業的重要課題。其中，應用捕食性天敵進行生物防治已被視為替代化學農藥的可行方案。然而，有效的生物防治需仰賴釋放足夠數量的捕食性昆蟲，使得其大規模飼養成為實現生物防治的一大挑戰。以常作為防治材料的瓢蟲為例，雖多以成蟲釋放，但其幼蟲階段的飼養卻佔據主要生產成本。 為了因應大規模飼養的需求，本研究設計並開發一套整合式自動化餵食系統，應用於小十三星瓢蟲(Harmonia dimidiata)幼蟲之獨居式飼養流程，以降低飼養流程中所需之人力與飼料成本。該系統硬體使用高精度之滾珠螺桿建立工作平台，並搭載以振動馬達與螺桿-活塞為驅動機構的天然與人工飼料餵食模組，能精準定位並完成餵食動作，在定位餵食的準確率上達100%，餵食量亦經由Bayesion假設檢定證實具高度一致性(Bayesian Factor ≪ 1)。此外，系統整合影像辨識功能，利用YOLOv8模型辨識幼蟲齡期與輪廓分割，盲測階段之分割成功率達99.44%。透過分割結果可提取六項幾何特徵，進而建立生產曲線，以評估幼蟲生長趨勢。在人工飼料方面，研究比較不同蛾卵比例之飼料配方，結果顯示人工飼料添加30%地中海粉螟蛾卵之組別，不僅發育時間與純地中海粉螟蛾卵組之差異低於兩天，存活率更達87.50%，並具成本優勢(僅為純蛾卵餵食之90.22%)。最終，建構餵食決策演算法，修正不合理之齡期預測，並排除發育遲緩之個體，以提升飼料利用效率。該演算法於餵食次數與飼料選擇之正確率分別為98.70%與99.03%，整體餵食正確率達97.74%。本研究所開發之系統提供了自動化且智慧化的昆蟲生產技術，可有助於提升飼養的規模。後續可持續優化其影像辨識模型與餵食決策，以提升其泛用性。未來可望應用此系統架構於其他昆蟲物種，以此作為昆蟲生產規模化且標準化之工具。

Although Pesticides have greatly enhanced global food production, the associated environmental and health concerns have increasingly drawn attention from both the public and academia. As a result, the development of environmentally friendly and sustainable pest management strategies has become a critical issue in modern agriculture. Among these, the use of predatory natural enemies for biological control has emerged as a viable alternative to chemical pesticides. However, effective biological control relies on the release of sufficient numbers of predatory insects, making their large-scale rearing a key challenge. In the case of ladybird beetles, although adults are typically released in the field, the rearing of larvae accounts for the majority of production costs. To meet the demands of large-scale rearing, this study developed an integrated automated feeding system for the individual rearing of Harmonia dimidiata larvae, aiming to reduce labor and feed costs. The system incorporates a high precision ball screw-driven platform equipped with two feeding modules: a vibration motor-based natural feed dispenser and a screw-piston-based artificial feed dispenser. The system achieves 100% accuracy in positioning and feeding, and Bayesian hypothesis testing confirmed high consistency in the amount of feed delivered (Bayesian Factor ≪ 1). A machine vision module system employing YOLOv8 was developed for larval instar recognition and contour segmentation, achieving a segmentation accuracy of 99.44% in blind tests. From the segmented contours, six geometric features were extracted to construct larval growth curves, enabling the assessment of developmental trends. In terms of artificial diet, experimental results showed that artificial diet with 30% Ephestia kuehniella eggs shortened development time to within two days of the pure egg group, achieved an 87.5% survival rate, and reduced feeding cost to 90.22% of the pure egg diet. A feeding decision algorithm was also developed to correct unreasonable instar predictions and eliminate underdeveloped individuals, thereby improving feed efficiency. The algorithm achieved 98.70% accuracy in feeding decision and 99.03% accuracy in feed selection, with an overall feeding accuracy of 97.74%. The system developed in this study offers an automated and intelligent approach to insect production, supporting scalable and standardized production. Future work may focus on enhancing the versatility of the system through improvements in image recognition models and feeding decision algorithms. This framework could also be adapted for other insect species and serve as a foundational tool for the standardization and industrialization of insect rearing.