基於影像分析追踪番茄生長狀態-以傳統番茄溫室為例

近年來，隨著人工智慧的興起，政府開始推動智慧農業，雖然在環境與植物生理感測上有著很好的發展，但卻較少將影像辨識的技術結合到農業上。本次研究將通過每日追踪埔里鎮育英國小傳統溫室內種植的番茄圖片色塊佔比的變動並結合使用TensorFlow平台，運用現行Object Detection演算法Faster R-CNN建立深度學習模型辨識每日番茄生長的特徵。研究由番茄的苗期開始追踪，並在採收期結束，共78天。研究將對比兩種辨識方式與番茄實際生長情況的差異，以檢視兩種方式的準確率。通過研究發現，兩種辨識在不同的生育期中有著不同優勢與劣勢。追踪色塊面積的方式在辨識綠色面積時基本符合番茄生長的情況，黃色與紅色容易受到外在因素的影響導致其準確度不高。而物件辨識的方式並不適合使用用來辨識番茄的生長情況，但在辨識花朵與果實時較追踪色塊面積的方式更準確。

關鍵字

番茄；傳統溫室；影像辨識； TensorFlow ； Faster R-CNN

並列摘要

Recently, with the rise of artificial intelligence, the government has also promoted and encouraged smart agriculture. Despite having great development in environmental and plant physiological sensing, the use of image sensing technology is still rare in agricultural field. This study is based on daily image tracking to monitor the change in the proportion of color patches of tomato in YuYing Elementary school's traditional greenhouses. It also uses Faster R-CNN to establish a deep learning model to investigate the characteristics and process of tomato’s daily growth. The study began with the seedling stage of the tomato and ended at the end of the harvest stage for a total of 78 days. This study shown the difference between the two identification methods and the actual growth of the tomato to examine the accuracy of the two methods. This study will track the change in the proportion of tomato photoblocks grown in the YuYing Elementary school's traditional greenhouses in Puli Township. TensorFlow and Faster RCNN were also used to identify the characteristics of daily tomato growth to supplement the deficiency. The study began with the seedling stage of the tomato and ended at the harvesting stage. The total duration needed is around 78 days. The study has shown the difference between two identification methods. The actual growth of the tomato represents the accuracy of the two methods. Through research, it is found that both types of identification have different advantages and disadvantages in different growth stages. The way to track the area of the color patch is basically consistent with the growth of the tomato when identifying the green area. The yellow and red are susceptible to external factors and the accuracy is not high. The study shown that object detection is more accurate in identifying flowers and fruits compared to that of tracking the area of patches.