應用YOLOv8-ByteTrack與Transformer於巢內蜜蜂行為之影像分析

Abstract

蜜蜂(Apis mellifera)在農作物中扮演重要的授粉者角色，對農產業具有重要影響力。然而近年來，蜜蜂數量大幅減少，對農業產量造成嚴重威脅。因此，了解影響蜜蜂的習性、行為和生長因素等，對於幫助蜜蜂族群永續生存至關重要。先前研究多著重於監測巢外蜜蜂行為，或僅針對巢內少數個體進行觀測與分析，對整體巢內活動的系統性探討仍相對不足。本研究旨在結合影像辨識技術與多目標追蹤演算法，並應用無監督與監督式機器學習進行比較分析，探索蜜蜂巢內行為之模式及巢內分工與區域之對應關係。檢測與追蹤方面，研究比較YOLOv8五種變體表現，並選擇表現最優的YOLOv8x（mAP50 = 98.84%，Precision = 97.29%）與ByteTrack MOT結合，確保蜜蜂個體追蹤穩定性（MOTA = 89.34%，IDF1 = 93.41%）。接續進行兩種特徵學習策略：一是無監督式方法，採用Transformer-based Autoencoder對連續時序下的軌跡進行特徵擷取與編碼，並以HDBSCAN聚類演算法進行行為分群與趨勢分析，PCC與ICR達90%，錯誤率降低80%；二是監督式方法，以觀察紀錄為基礎進行蜜蜂爬行軌跡特徵分析與真實行為對應，並使用Transformer架構同樣對連續時序下的軌跡進行訓練與分析，各項指標均超過85%。為了驗證分析結果的準確性與實用性，最後以人工紀錄之巢房功能分區作為驗證與比較依據，評估兩種模型的分析結果。結果顯示，無監督模型適合探索大規模運動動態，監督模型則有助於解析行為與巢區的關聯。總體而言，本研究證明了蜂箱內行為分析的可行性，為非侵入式、即時監測系統奠定基礎，有助智慧養蜂與生態永續。無監督模型捕捉集體運動趨勢的能力，顯示其在蜂群崩解症（CCD）早期預警上具有潛力，而監督模型若進一步研究與開發，將可實現更高效且友善的蜂群管理系統，兩種模型各有優缺，為蜜蜂生態研究提供支持。

Honey bees (Apis mellifera) are vital pollinators in agriculture, yet recent population declines threaten crop yields. Understanding in-hive behavior and factors affecting colony sustainability is therefore critical. Previous studies have largely focused on foraging or limited in-hive observations, leaving overall colony dynamics underexplored. This study integrates computer vision with multi-object tracking (MOT) and compares unsupervised and supervised machine learning approaches to analyze in-hive behavior and the correspondence between behavioral roles and hive regions. YOLOv8x (mAP50 = 98.84%, Precision = 97.29%) combined with ByteTrack (MOTA = 89.34%, IDF1 = 93.41%) enabled reliable individual tracking. The unsupervised framework used a Transformer-based autoencoder with HDBSCAN clustering to extract and group trajectory features, achieving 90% on PCC and ICR with an 80% error reduction. The supervised framework trained a Transformer classifier on annotated data, achieving over 85% across all metrics. Analyses were validated against manually annotated hive functional regions. Results shows that the unsupervised model effectively captures large-scale movement patterns, whereas the supervised model clarifies behavior-region associations. Overall, this study demonstrates the feasibility of in-hive behavior monitoring, providing a foundation for non-invasive, real-time systems that support smart beekeeping and ecological sustainability. The unsupervised approach may serve as an early-warning indicator for Colony Collapse Disorder (CCD), while supervised models could enhance colony management efficiency, highlighting complementary strengths for bee ecological research.