- 學位論文
基於影像處理和深度學習的蜜蜂蜂巢智能監測和預測系統
A Smart Monitoring and Forecasting System for Honey Bee Colony Based on Image Processing and Deep Learning
摘要
本研究提出一個蜜蜂即時影像監測系統，藉由計算蜜蜂進出蜂箱的數量來監測其活動。影像會在蜂巢的入口處進行拍攝，並以兩種不同的方法將影像中的蜜蜂擷取並檢測：背景相減法與深度學習。以Kalman Filter和Hungarian algorithm可以實現對蜜蜂移動軌跡的追蹤。追蹤演算法用以計算單隻蜜蜂的進出活動，與人工計數相比，準確率達93.9±1.1%。蜜蜂的進出活動與感測器量測到的環境參數會藉由4G無線路由器上傳至遠端伺服器，以進行後續分析。在物體追蹤演算法下，訓練出一個輕量型的即時物體檢測與基於深度學習的分類模型，可以計算並將蜜蜂辨識為攜帶花粉或沒有攜帶花粉兩類。辨識蜜蜂是否攜帶花粉的F1-score為0.94，precision與recall分別為0.91與0.99。關於計算蜂群採食行為演算法的平均誤差百分比，攜帶花粉的蜜蜂數量為8.45±2.72%，而進入蜂箱的蜜蜂總數為10.55±2.10%。為測試影像監測系統於連續監測蜜蜂採食行為之性能、研究天氣因素造成的影響，及評估農藥對蜜蜂進出行為之影響，本研究進行了些許實驗。記錄進出蜂箱的蜜蜂數量，並根據每小時和每日紀錄的數量來計算，以進行下一步分析。實驗結果與分析表明，每日花粉採集的比例为24.5±3.5%；單個蜂箱每日採集的花粉量約49.1±11.0g。採集的花粉量隨溫度與光照增加而增加，隨相對濕度、雨量和風速增加而減少。採集花粉的行為在大雨或微風情況下有顯著減少。應用數據分析方法從已知的蜜蜂進出數量中提取信息。對蜂群損失率的預測模型，可以成為管理蜂群健康的一個重要工具，並能為早期預警方法提供條件，以瞭解影響蜜蜂群體的潛在異常狀況。預測模型是以TCN (temporal convolutional neural networks)模型為基礎作訓練，用以預測隔日的蜂群損失率。通過對特徵進行重要性分析、特徵選擇與超參數對預測模型進行優化。預警級別的訂定，是以isolation forest演算法將蜂群損失率分為正常或異常。綜合演算法的兩個蜂群損失率測試集，源自於同個養蜂場的多個蜂箱。測試結果表明，預測模型的WMAPE (weighted mean average percentage error)為17.1±1.6%，而預警模型的準確率為90.0±8.5%。 本研究開發的預測模型可促進對蜂群的有效管理，並防止蜂群崩潰。該自動影像監測系統可作為一種高效可靠的工具，供研究者深入瞭解蜜蜂的採食行為，並協助蜂農管理蜂箱。
並列摘要
This research presents a real-time imaging system for monitoring honey bee activity by counting the honey bees entering and exiting the beehive. Images of the honey bees are continuously acquired at the beehive entrance. The honey bees in images are detected using two different approaches: background subtraction and deep learning based. Tracking of honey bees is achieved using an integrated Kalman Filter and Hungarian algorithm. The tracking algorithm is used to determine the incoming and outgoing activity of individual honey bees and has an automatic counting accuracy of 93.91.1% in comparison with manual counting. The in-and-out activity and environmental information from sensors are transmitted to a remote server via a 4G LTE router for later analyses. A lightweight, real-time object detection and deep learning-based classification model, supported by an object tracking algorithm, is trained for counting and recognizing honey bees, separating them into pollen or non-pollen bearing classes. The F1-score is 0.94 for pollen and non-pollen bearing honey bee recognition, and the precision and recall values are 0.91 and 0.99, respectively. For the foraging trip counting algorithm, the mean average percent errors of the pollen bearing honey bee count and the total incoming honey bee count are 8.45±2.72% and 10.55±2.10%, respectively. Several experiments are performed to test the performance of the imaging system in continuous monitoring of honey bee pollen foraging behavior as well as to investigate the effect caused by weather factors and assess the pesticide effect on the in-and-out behavior. The incoming and outgoing honey bee counts were used to calculate indices based on the hourly and daily counts to assess pesticide effect on the foraging behavior. The experimental results and analyses revealed that the daily pollen foraging trip ratio was 24.5±3.5%; a single beehive collected about 49.1±11.0 g of pollen per day. The pollen foraging trip count increased with increasing temperature and light intensity, and decreased with increasing relative humidity, rain level and wind speed. A significant reduction of pollen foraging activities was observed in heavy rainfall or gentle breeze conditions. Data analytics of the in-and-out counts contributed to the development of the forecasting model, the assessment of the pesticide effect and weather effect on the foraging behavior. This forecasting model for the population loss rate of a honey bee colony can be an essential tool in honey bee health management, a component of early warning methods of potential abnormalities affecting a honey bee colony. The forecasting model is based on temporal convolutional networks (TCN) to predict the following day’s population loss rate. The forecasting model was optimized by conducting feature importance analysis, feature selection, and hyperparameter optimization. Using an isolation forest algorithm, the population daily loss rate is classified as normal or abnormal. The forecasting model together with an isolation forest algorithm was tested on two population loss rate datasets collected from multiple honey bee colonies in a honey bee farm. The test results show that the forecasting model can achieve a weighted mean average percentage error (WMAPE) of 17.1±1.6%, while the warning level determination method reached 90.0±8.5% accuracy. The forecasting model developed through this research can be used to facilitate efficient management of honey bee colonies and prevent colony collapse. The automated imaging system can be applied as an efficient and reliable tool for researchers to gain deeper insights into honey bee foraging behavior, and help beekeepers achieve better beehive management.