請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83219
標題: | 以無人機進行坡地茶園因蟲害所導致異常的即時辨識研究 Research on Onboard UAV Real Time Pest Affected Anomaly Detection for Slope Tea Farm |
其他標題: | Research on Onboard UAV Real Time Pest Affected Anomaly Detection for Slope Tea Farm |
作者: | 張弘熙 Hung-Hsi Chang |
指導教授: | 張時中 Shi-Chung Chang |
關鍵字: | 茶園,巡檢自動化,無人機,多光譜影像,蟲害導致的異常偵測,機上邊緣運算, Tea Farm,Automated Inspection,UAV,Multispectral Image,Pest Affected Anomaly Detection,Onboard Edge Computing, |
出版年 : | 2022 |
學位: | 碩士 |
摘要: | 在台灣約有12,196公頃的茶園,一半為平地,一半為坡地。種植茶葉每年有60億新台幣的產值。在茶葉種植的過程中,透過巡視茶園依生長狀況處理栽植工作,保持茶葉在良好的健康狀態是茶農重要的工作。茶樹的異常包括缺乏營養、水分、受蟲害和受病害。由於從業人口的老化與數量衰退,使用無人機替代進行巡檢已是趨勢。無人機巡檢目前以多光譜儀從空中拍攝目標茶園。多光譜影像辨識前處理流程透過將在不同位置拍攝的多光譜影像拼接以校正拍攝時環境光線影。此過程耗時一至數天,無法如人力巡檢一樣在巡檢結束當下就知道生長情形並且馬上處理。若巡檢無人機上能即時進行單張未拼接多光譜影像處理辨識,使無人機降落前就辨識完整片目標茶園,可當下採取進一步的巡檢或拍攝處置,蒐集更精細的資訊。 基於上述需求,本研究專注於開發坡地茶園無人機上即時多光譜影像辨識指標與方法,以蟲害所導致異常(以下簡稱異常)為目標,並探討在坡地茶園進行維持相對高度取像略有差異時的影像與處理方法。主要研究問題、相應的挑戰和新提出並設計解決方案為: P1 坡地茶園有標記之多光譜影像蒐集問題: 如何定義和蒐集標記?如何產生有標記之茶園? C1 尋找適合且可供實驗茶園以及產生實驗所需要的健康/不健康茶樹。 M1 於茶業改良場文山分場坡地茶園,透過噴灑與未噴灑農藥產生病蟲害的實驗組與對照組區域,以人工觀察計算取樣範圍內受蟲害影響茶芽數與茶芽總數比例蒐集病受蟲害區茶樹實際的異常比例作為電腦辨識的標記。 P2 因蟲害所導致的異常辨識問題: 2.1 光譜特徵選擇子問題:要選擇哪些植生指數作為光譜特徵以分辨茶樹和非茶樹以及判斷因蟲害所導致的異常程度? 2.2 空間解析度差異所造成的辨識子問題:不同的空間解析度會對蟲害所導致的異常辨識造成什麼影響? C2.1 茶樹及雜草皆為綠色植物,使用目前文獻中一般用於突顯植物生長的植生指數無法明確辨別差異。蟲害部分文獻中採用植生指數繁多,可突顯的特徵皆相異,但無文獻探討與受蟲害茶樹相關的植生指數。 C2.2 若異常辨識會受到空間解析度影響,找出如何影響是挑戰。若異常辨識不會受到空間解析度影響,找出原因是挑戰。 M2.1 經光譜分析發現茶樹在綠色頻段的亮度較雜草暗,在紅邊頻段則相反。搜尋後發現GNDRE指數計算公式包含紅邊與綠色亮度相減,適合用來突顯差異,茶樹辨識達到97.56%準確率、異常比例預測達到0.0753 RMSE及0.6946 R2。 M2.2 辨識不同空間解析度之相同區域影像後發現異常辨識結果有差異。透過主成分分析將不同空間解析度的樣本投影後分析,取樣區域實際大小越小、空間解析度越高時,特徵分布越分散,容易產生極端辨識結果。 P3 即時異常辨識演算法設計問題:如何設計異常辨識演算法以達到即時? C3 影像校正耗時6秒,因高度差異所需的各頻段影像對齊耗時10分鐘。茶園巡檢對即時異常辨識的要求為在影像重疊率小於0之前完成辨識,在相對高度10公尺、秒速0.8公尺飛行時需在7.6秒內辨識完畢以達到即時,如何減少前處理時間是挑戰。 M3 因相機校正參數不隨時間變化,透過使用預先以參數產生並儲存的校正檔案節省3秒計算時間。無人機高度維持在10公尺時雖有10%誤差,但對齊的平移量最大差異僅有2.4 pixels。以一組對齊參數套用至所有多光譜影像可省下10分鐘運算時間。平均執行時間最終減少至5.6秒,在相對高度10公尺以秒速0.8公尺飛行時可以達到即時。 本論文的研究發現、貢獻和價值如下: 1. 相較現有以近距離拍茶葉攝病徵使用空間特徵辨識,本研究透過空拍之未拼接單張茶園多光譜影像以光譜特徵辨識蟲害危害程度。 2. 設計坡地茶園巡檢機載即時異常辨識演算法,執行時間5.6秒,在相對高度10公尺以秒速0.8公尺飛行時達到全區域涵蓋。 3. 與現行無人機茶園巡檢相比,可提前數小時至數天提供巡檢異常辨識結果。 4. 無人機相對地面高度對於茶園巡檢即時異常辨識的重要性。 5. 機載即時異常辨識之反饋具提供無人機飛行中即時降低高度聚焦以及調整飛行路徑的發展潛力基礎。 In Taiwan, there were 12,196 hectares tea farms in year 2020. About half of the tea farming area is situated above plains, and the other half is located in mountainous areas. The value produced by growing tea is approx. 6 billion NTD per year. During tea plants growing, inspections conducted by tea farmers are important to make sure the tea plants are in good health condition. The anomalies of tea plants include lacking of nutrients, dehydration, diseases and pests. Due to the existing tea farmers getting older, using UAV to replace tea farmers for inspection seems to be a promising solution. Nowadays, inspection by UAV uses multispectral camera to image target tea farm from the air. The preprocessing of multispectral images (MSIs) stitches the images taken at different positions to calibrate the brightness. The process takes several days. It is not able to produce inspection result right after inspection as manual inspection for tea farmers to address immediately. If the taken MSI can be detected onboard the inspection UAV without stitching, further actions such as imaging closer for finer details could be carried out. Based on the demands mentioned above, this research aims at developing a method for onboard UAV real time MSI pest affected anomaly detection for slope tea farm. Pest affected anomaly, hereinafter referred to as anomaly. We also discuss how to deal with the slight differences caused by error of UAV maintaining its height. The problems, corresponding challenges and proposed designs are as following: P1 Slope tea farm labeled MSIs collection problem: How to define label for anomaly? How to generate tea plants with labels and how to collect? C1 Finding a tea farm and generating healthy and pest affected regions. M1 At a slope tea farm at Tea Research and Extension Station-Wenshen Substation, we generated experimental group and control group regions of pests and diseases by spreading pesticides. We calculated the ratio between the number of pest affected new shoots and the number of new shoots in sampling area by visual observation. We collected the actual anomaly ratios in pest affected region as labels for computer detection. P2 Pest affected anomaly detection problem: 2.1 Spectral features selection subproblem: Which VIs should be selected as spectral features to distinguish tea and non-tea plants and to detect anomaly? 2.2 Sensitivity of anomaly detection model to difference in spatial resolutions subproblem: How different spatial resolutions affect the anomaly detection algorithm? C2.1 Both of tea plants and weeds are green plants and share similar spectrum. The VIs for indicating plant growth are not able to distinguish clearly. For pest affected anomaly, there were numerous related VIs from literatures. No literature related to the VI of pest affected tea plant was found. C2.2 If anomaly detection algorithm is affected by different spatial resolutions, to find out how is a challenge. Or if anomaly detection algorithm is not affected by different spatial resolutions, then to find out why is a challenge. M2.1 We found that the green color of tea plants is darker than that of weeds and the RE band shows the opposite. The GNDRE index contains RE minus green term, which is suitable to amplify the difference. 97.56% 0/1 accuracy was achieved for tea classification. 0.0753 RMSE and 0.6946 R2 were achieved for anomaly ratio prediction. M2.2 The detection results from MSIs of the same place with different spatial resolutions were different. Principal component analysis was used to project samples of different spatial resolutions for analysis. For smaller real world sampling size, i.e., higher spatial resolution, the distribution was more scattered, thus noisier, and may produce extreme predictions. P3 Real time anomaly detection algorithm design problem: How to design anomaly detection algorithm that can achieve real time? C3 Image calibration took 6 seconds. Band image alignment for height maintaining error took 10 minutes. The requirement for tea farm real time anomaly detection is that the algorithm produces result before overlap ratio smaller than 0. At 10 meters height and 0.8 m/s speed, it needs to finish detection in 7.6 seconds to achieve real time. To make the preprocessing as fast as possible is challenge. M3 The camera calibration parameters are time invariant, thus our algorithm uses the stored profiles pregenerated from parameters for reducing 3 seconds. Height maintaining at 10 meters had 10% error for our UAV. However, the maximum difference in translation of alignment was only 2.4 pixels. By applying one alignment parameter to all MSI reduced computation time by 10 minutes. The average running time was 5.6 seconds ultimately and it achieves real time when flying at 10 meters height and 0.8m/s. The findings, contributions and values are as following: 1. Compared to imaging the pattern on tea leaves caused by pests in close distance and detecting based on spatial features, this research used non-stitched individual MSI of tea farm to detect pest affected anomaly by spectral features. 2. Designed a real time anomaly detection algorithm for onboard UAV tea farm inspection with to 5.6 seconds running time covering full target area at 10 meters height and 0.8 m/s speed 3. Provides anomaly detection results hours to days advance than current tea farm inspection practice 4. Importance of maintaining height for UAV while performing tea farm inspection with real time anomaly detection algorithm 5. The feedback from the onboard real time anomaly detection provides the potential to lower the height for focusing on anomaly and perform onboard dynamic path planning. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83219 |
DOI: | 10.6342/NTU202204009 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 電機工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-2509202217102400.pdf | 5.5 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。