應用深度學習演算法於茶芽生長監測與關鍵期預測

Abstract

茶葉是臺灣重要的農產品之一，飲料市場的蓬勃發展更促進茶產業的持續成長。農產業近年來面臨嚴重缺工的困境，茶產業亦深受衝擊。結合感測器、物聯網、影像判別的智慧農業技術開發，可協助農民更便利及有效率地管理場域及生長狀況。對茶農來說，若可即時監測茶樹生長狀態即可減少巡園人力；而準確地判斷重要的生長期及可能的採收時機，更對給水、給肥、防治管理，或採收人力安排更是大有幫助。因此本研究旨在開發茶芽生長關鍵期預測模型，結合田間監測模組以蒐集茶樹影像及環境資訊，並利用深度學習模型識別茶芽數量。透過生長預測模型來推估未來茶樹生長趨勢，以判別茶樹的生長階段，快速生長期及最佳採收期。 本研究主要分為三個部分組成：田間IoT監測模組、茶芽識別模型及茶芽生長預測模型。首先，IoT感測模組由溫濕度感測器、Raspberry Pi 和Raspberry Pi camera組成。透過模組蒐集 2020 年至 2023 年桃園、南投、屏東等三區域茶園中，臺茶12號、臺茶18號與青心烏龍三品種之生長影像約3000餘張。接續則應用深度學習方法開發茶芽識別模型，辨識影像中之茶芽數量。透過取得多組完整茶季之茶樹生長數據，搭配茶芽識別結果，與在地氣象站預報資訊，即可採用時間序列模型建立茶芽生長預測模型。利用生長曲線中的趨勢特徵，以Gompertz模型擬合曲線後，則可進一步預測關鍵的快速生長日期與最佳採收日期。 試驗結果顯示，茶芽識別模型以YOLOv9 (You Only Look Once v9)架構於驗證集mAP達到0.878，於測試集達到0.562較佳。針對六季不同品種茶樹進行茶芽識別與計數，預測與實際值間數量之平均絕對誤差分別為春茶 6.55、夏茶 5.99，與秋茶7.89。此結果在數量的誤差少於10，已具有可取代的應用潛力。茶芽生長預測模型則以LSTM-FCN (Long Short-Term Memory Fully Convolutional Network) 表現最佳，於三季測試中，快速生長期及最佳採收期之預測誤差均在± 2天內。透過茶芽生長模型預測出的未來生長趨勢，亦可瞭解採收時的茶芽數量，於三季測試中，預測與實際數量的誤差皆在10個茶芽內，此結果可為日後農民的採收作業提供產量評估參考。本研究結果成功基於YOLOv9模型架構建立茶芽識別與計數模型，並透過識別出的茶芽數量建立生長預測模型，實現茶樹生長關鍵期預測。同時該模型也提供產量預測的客觀參考，期望有助於提高茶農的管理效率和收益。

Tea is one of Taiwan’s essential agricultural products, and the booming development of the beverage market has promoted the continued growth of the tea industry. However, the agricultural industry has faced serious labor shortages in recent years, significantly affecting the tea industry. The development of smart agricultural technology that integrates sensors, the Internet of Things, and image recognition can help farmers manage fields and monitor growth conditions more conveniently and efficiently. For tea farmers, real-time monitoring of tea growth can reduce the labor required for patrolling the plantation. Accurately identifying critical growth periods and potential harvest times can also improve arrangements for water supply, fertilization, pest control, and harvesting labor. This study aims to develop an auxiliary system combining field monitoring modules to collect tea images and environmental information, using a deep learning model to identify the number of tea buds. The growth predictive model estimates the future growth trend of tea to determine their growth stage, rapid growth period, and optimal harvest day. The study is divided into three parts: the field IoT monitoring module, the tea bud detection model, and the tea bud growth predictive model. The IoT sensing module includes a temperature and humidity sensor, a Raspberry Pi, and a Raspberry Pi camera. This module collected more than 3,000 images of the growth of TTES No. 12, TTES No. 18, and Chin-Shin Oolong from tea plantations in Taoyuan, Nantou, and Pingtung from 2020 to 2023. The deep learning method developed a tea bud recognition model to identify the number of tea buds in the images. By obtaining multiple sets of tea growth data for complete tea seasons and combining these with tea bud identification results and local weather station forecasts, a time series model was used to establish a tea bud growth predictive model. The trend characteristics in the growth curve were analyzed and fitted with reciprocal exponential and Gompertz models to predict critical rapid growth period and optimal harvest days. The experimental results of this study demonstrate that the tea bud detection model, utilizing the YOLOv9 (You Only Look Once v9) architecture, achieved a mAP of 0.878 on the verification set and 0.562 on the testing set. For tea bud detection and counting of different tea varieties in six seasons, the mean absolute error (MAE) between the predicted and actual tea bud counts was 6.55 for spring tea, 5.99 for summer tea, and 7.89 for autumn tea. With a count error of less than 10, this model demonstrates potential for practical application. The tea bud growth predictive model performed best with LSTM-FCN (Long Short-Term Memory Fully Convolutional Network). In the three-season test, the prediction errors for the rapid growth day and the optimal harvest day were within ± 2 days. The future growth trend predicted by the growth predictive model also closely matched the actual number of tea buds at harvest, with an error of within 10 tea buds in the three-season test. This result can guide farmers’ harvesting operations and provide a reference for yield evaluation. This study successfully established a tea bud identification and counting model based on the YOLOv9 model architecture and a growth prediction model based on the identified tea bud counts. The models accurately predicted the critical periods of tea tree growth and provided an objective reference for yield prediction, potentially improving the management efficiency and profitability of tea farmers.