以卷積神經網路分析方法應用於除草系統之研究

Abstract

雜草在作物栽培體系扮演著不能忽視的角色，因為它對陽光、水份、養份、空間等資源的競爭，會直接影響到作物的生長。雜草的防治若以化學方法處理，會污染環境及農產品，而以人力方法處理，既費時又費工，效率非常低。近年來研究有關於機械除草越來越受到關注，機械除草相比化學除草和人工除草具有污染少、效率高的優點，採用機械除草設備進行田間除草，以目前研發之設備而言，大部分只做到行間的除草，株間的除草仍需靠人工。甘藍於種植前期，其生長勢與周遭雜草相當，需要進行密集的除草，在缺工的農業生產模式下，造成農民許多困擾。 本研究應用影像定位技術開發智慧型田間甘藍除草系統，進行物理性之除草，使用攝影機擷取田間帶有雜草之甘藍影像照片474張，分為379張之訓練組及95張之預測組，透過卷積神經網路 (Convolutional Neural Network, CNN)的影像處理方法，取出特徵且分類來辨識甘藍菜和雜草。在驗證的圖片中共計有 381 株甘藍，只有 3 株沒被辨識出來，成功率為99.2%，而且沒有雜草被辨識成甘藍，透過框選出來的結果也可以得到甘藍植株的位置。以建立之模型進行實地測試，即使在雜草比訓練樣本更多的環境下，一樣能辨識出甘藍，共將49珠作物辨識成功，5未辨識出，將3珠雜草辨識成作物辨識率為86.0%。 進行實驗之後證實其能有效和準確的辨識作物之後進行除草爪迴轉以達到攪動土壤除草的功能，採用之三種轉速分別為 0、9、23 rpm，9、23 rpm可以因應不同車速所面對的不同情況進行除草，能有70%成功率透過旋轉除草爪進行除草，研究結果顯示以智慧型除草系統，可達到精確、快速地獲得作物的位置資訊，並進一步實現自動化精準除草，達成降低農作物的生產成本，提升其生產力，更能達到保護環境和提供消費者健康和安全農產品。

Weeds play an important, non-negligible role in crop cultivation because their competition for sunlight, moisture, nutrients, space and other resources directly affects the growth of crops. Application of chemicals for weed control will pollute the environment and agricultural products, while physical treatments is time-consuming and laborious, which leads to low efficiency. Recently, research on mechanical weeding has received more and more attention. Compared to chemical and manual techniques, mechanical weeding can be more efficient and less pollution. Conducting mechanical weeding in the field, between-row weeding is the best way current technologies can do. Intra-row weeding relies on labor force. Therefore, while the technology of mechanical weeding has reached quite a breakthrough, the productivity has been making no progress. The growth potential of cabbage can be smilar to that of weeds. Therefore, intensive weeding is highly required. The situation can be bothering to farmers in the agricultural production mode with labor shortages. This research intends to develop an intelligent vegetable intra-row weeding system using image positioning technology to conduct physical weeding. Total of 474 cabbage images with weeds were captured in the field with camera, in which 379 of these images were used as training data, and the other 95 images were used as testing data. Through the image processing method of Convolutional Neural Network (CNN), the features were extracted and classified between identify cabbages and weeds. There were 381 cabbages in the verified images in total, only 3 of which were unidentified, with a success rate of 99.2%. No weed was identified as cabbage, and the positions of cabbages were also obtained. Conducting field tests with this model, we are able to identify cabbages in an environment where the number of weeds is far more than training samples. In the test, 49 crops were successfully identified, 5 were not. 3 weeds were wrongly identified as crops. The identification rate reached 86%. Tests show that after identifying crops from weeds, applying weeding rotors is able to remove weeds by rotating and stirring the soil. With the rotating speed of 0, 9, 23 rpm respectively, rotors are able to weed under different circumstances. The success rate of applying weeding rotors to weed is 70%. The research shows that applying smart weeding system allows people to locate crops and remove weeds automatically and precisely. The system is able to reduce the producing cost of agriculture and enhance the productivity. Furthermore, it protects the environment and provides consumers with healthy and safe agricultural products.