應用深度學習解析蜂群進出巢行為與花粉動態

Abstract

西方蜜蜂（Apis mellifera）在設施農業中是最主要的授粉者，約貢獻全球八成的農業授粉量。然而蜂群健康的好壞很大的決定了蜜蜂授粉的效果，尤其在食物較於單一的設施農業當中，農夫需要花費較多的資源去觀察蜜蜂。因此，本研究提出一套非侵入式影像監測系統，同時追蹤蜜蜂進出蜂巢行為以及收集蜜蜂攜帶花粉回巢之數據。(1) 利用 YOLO11 物件偵測模型偵測蜜蜂位置、大小與類別資訊。而 YOLO11 模型於兩種蜜蜂的偵測上達到了 90.08% 的 mAP@50-95。(2) 利用 PP-LiteSeg 語意分割模型分割影片中蜜蜂後腳上的花粉，並計算花粉面積。而PP-LiteSeg 模型在花粉分割方面的表現達到 61.65% 的 IoU。(3) 結合前兩個模型的結果、ByteSort 與計數演算法，獲得蜜蜂進出巢數量與蜜蜂帶回巢之花粉總面積。並且得出花粉面積與花粉重量之相關性。最後，本研究將監測系統應用於台灣南部的屏東縣萬巒鄉的百香果設施農業中，分析在有無花粉篩孔、蜂王與幼蟲的情況下，蜜蜂進出巢數量與採集花粉的積極度變化，並提供更進一步的蜜蜂活動與覓食的相關資訊，協助農民調整設施內蜂群的授粉與健康狀況。

The western honeybee (Apis mellifera) serves as the main pollinator in protected agriculture, contributing almost 80\\% of global agricultural pollination. However, colony health remains a critical determinant of pollination effectiveness. This issue is particularly pronounced in controlled environments such as greenhouse agriculture, where floral diversity is limited. Under such conditions, farmers must allocate additional resources to monitor honeybee activity. To address this challenge, this study introduces a non-invasive, image-based monitoring system that concurrently tracks entry and exit activities at the hive entrance and quantifies pollen loads carried by returning bees. (1) A YOLO11 object detection model was employed to identify the position, size, and category of individual bees, achieving an mAP@50-95 of 90.08% for two bee categories. (2) A PP-LiteSeg semantic segmentation model was used to segment hind leg pollen from honeybees in video frames, generating an IoU of 61.65% for pollen segmentation. (3) The outputs of both models were integrated using ByteSort and a custom counting algorithm to estimate the entry and exit numbers of the honeybees and the total pollen area carried back to the beehive. Subsequently, a correlation was established between pollen area and pollen weight. The correlation coefficient was 0.9735, indicating a strong positive correlation.Finally, the system was deployed in a passion fruit greenhouse in Wanluan Township, Pingtung County, southern Taiwan, in order to evaluate variations in foraging activity under different colony conditions, including the presence or absence of pollen trap, a queen bee, and larvae. The system provides valuable information on the foraging behavior of honeybees and colony performance within facility agricultural systems, thereby assisting farmers in managing pollination efficiency and colony health.