智慧辨識技術應用於甘藍除草整合系統之研究

Abstract

雜草防治是農業生產相當重要的一環，目前台灣多以除草劑作為雜草控制的手段。然而，除草劑對人體與環境皆有害。因此有機農業在近幾年備受推崇。現行有機農業大部分以人工進行除草，不但費時且人力成本高。本研究開發智慧除草系統，以智慧影像辨識結合機械除草的方法，希望可改善現行除草之方式。 本研究採用甘藍為作物對象，以CNN (Convolutional Neural Network, 卷積神經網路)的人工智慧分析法，對一到四週株齡甘藍苗進行訓練。樣本數共1828張照片(7432株甘藍)，校正組為1463張照片(5854株)，驗證組為365張照片（1578株甘藍），其建立之甘藍智慧辨識系統，對驗證組甘藍之召回率99.6%、精確率為100%及F1-Score為99.8%；以辨識模型為基礎所建立的影像追蹤模型，對同時期拍攝的三週株齡甘藍72株影像追蹤率為100%。由於曳引機缺少精確的車速偵測設備，且在田間行駛可能會遭遇打滑的問題導致誤差，因此本研究設計以影像追蹤法進行車速量測，並與標準測量值進行比較分析，建立之模式r2為0.9999，標準誤差SEC為0.11 cm/s，可用於本智慧除草系統車速之檢測。 除草機構設計可同時進行溝底、行間及株間除草。在行間及溝底的部分採用可移動高低位置之除草犁；株間除草則整合人工智慧影像辨識之結果，利用比例閥調控油壓馬達的流量，使除草爪旋轉避讓甘藍達到株間除草之目的。因應田間地面高低不平，利用四連桿機構配合導輪，對除草機構於畦面高度進行微幅調整。本系統對甘藍植株除草成功率為96.3%，除草效率為0.096公頃/小時，為人工除草效率8.35倍，人工除草成本為本研究之系統成本3.82倍。除草面積為畦面的總面積的96.56%，研究結果顯示智慧型除草系統可大幅提升除草效率，並可解決農業缺工問題，提高農業生產效率及降低生產成本，並取代化學除草的方法，降低環境污染達到永續農業的目的。

Weed control is important for agricultural production, and herbicides are currently used to control weeds in Taiwan's agriculture. However, herbicides are harmful to the health and the environment. Therefore, organic farming has been highly appreciated in recent years. Most of the existing organic farming uses manual weeding, which is not only time consuming but also has high labor cost. In this study, we developed an intelligent weeding system, which is a combination of intelligent image recognition and mechanical weeding method, hoping to improve the current weeding practice. This research developed an intelligent weeding system combining the Convolutional Neural Network (CNN) image processing with weeding mechanisms. The total of 1828 pictures (total 7432 images of cabbage seedlings ) of 1-4 weeks old cabbages seedling were taken as the training samples, of which 1463 pictures (total 5854 images of cabbage seedlings) were selected as the calibration group, and 365 pictures (total 1578 images of cabbage seedlings) as the validation group. An intelligent identification system for cabbage seedlings was then established, and the has 99.6% recall rate, 100% accuracy rate and 99.8% F1-Score. The coefficient of determination (r2 value) of the vehicle speed measurements by image tracking method in comparison with the standard measurements is 0.9999, and SEC (Standard method of Calibration) is 0.11 cm/s, which can reduce the development of the hardware and achieve a more accurate speed detection. The weeding mechanism was designed to perform weeding operationd for the bottom of the ditch, inter-row and intra-row weeds.Weeding plows with movable high and low positions were used for inter-row and the ditch-bottom weeding. Intra-row weeding system integrated the intelligent image recognition, used a proportional valve to regulate the flow of the hydraulic motor, so that the weeding claws that advanced with the tractor and rotated to avoid hitting the cabbage seedlings and only conducting the intra-row weeding. The weeding success rate for cabbage seedlings can reach to 96.3%. The weeding efficiency is 0.096 hectares per hour, which is more than 8.35 times the manual weeding efficiency, and the manual weeding cost is more than 3.83 times compare to developed system. Including weeding between plants and between rows of furrows, the weeding area is 96.56%. The research results show that the intelligent weeding system can greatly improve the weeding efficiency, and can solve the problem of labor storage in agriculture, improve agricultural production efficiency and reduce production costs, replace chemical weeding methods, reduce environmental pollution, and achieve the goal of sustainable agriculture.