請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/87174
標題: | 應用深度學習方法於地理圖資系統水稻田坵塊航照圖之辨識與分期 Using Deep Learning Method to Conduct Paddy Field Detection and Staging in Aerial Images in Geographic Information System |
作者: | 周易昕 Yi-Shin Chou |
指導教授: | 周呈霙 Cheng-Ying Chou |
關鍵字: | 深度學習,實例分割,航照影像,水稻田,水稻物候期,植生指標, Deep learning,Instance segmentation,Aerial image,Paddy field,Rice phenological stage,Vegetation index, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 水稻作為台灣最重要的糧食作物之一,在農產預估及災損評估上時常有大範圍辨識的需求,然而台灣現階段對於水稻田的監測與管控作業,主要仰賴農試所專家們透過實地訪查、比對地籍圖資料,以及藉由衛星或航照影像輔助,再以人工的方式逐一辨別水稻的種植分佈,過程中除了要進行水稻田資料標記作業,還要定期更新、維護資料的正確性,使得整體作業流程曠日廢時,因此面對即時且大範圍的辨識需求往往欠缺效率。本研究以 ArcGIS Pro 軟體作為開發深度學習模型的工具,開發 Mask RCNN 實例分割模型辨識和分割航照影像中的水稻田。使用了 2018 年至 2019 年收集的航空影像,涵蓋彰化、雲林、嘉義和台南的水稻種植密集區。模型標記檔創建了四個類別,包括水稻生長、黃熟、收割階段和其他作物。另外研究在影像前處理階段對原始航照影像的波段資訊進行修改,利用不同的植生指標(NDVI、CMFI、DVI、RVI、GRVI)創建不同的訓練資料集,測試不同植生指標對於水稻分期辨識的表現。結果發現,以 ResNet-50 為特徵提取骨架的模型在 mAP 方面的表現整體優於 ResNet-101,其中 RGB + DVI、RGB + NIR 和 RGB + GRVI 影像資料集表現最好,mAP 分別為 74.01%、73.81% 和分別為 73.72%。而在水稻田分期辨識和分割的問題上,水稻生長階段建議使用 RGB + CMFI 影像訓練的模型,水稻黃熟階段建議使用 RGB + NIR 影像訓練的模型,水稻收割階段則建議使用 RGB + GRVI 影像訓練的模型,dice coefficient 分別為 79.59%、89.71% 和 87.94% 。模型分期辨識和分割結果可以提供水稻生長狀態的資訊,從而提升水稻生產管理的效率和準確性。該方法也可用於任何農作物的大規模檢測,減輕研究人員的負擔,提高土地利用調查的效率。 Given the significance of rice in Taiwan's agriculture, efficient methods for detecting and mapping paddy fields are critical for the effective management of agricultural production, prediction of yields, and assessment of the damage. Currently, researchers at the Taiwan Agricultural Research Institute use a combination of site surveys, cadastral maps, and satellite or aerial images to identify rice planting areas' distribution manually. However, maintaining the accuracy of the paddy field data by regularly updating the labels is time-consuming, especially when faced with the need for large-scale and immediate detection. This study aimed to detect and segment paddy fields in aerial images using Mask RCNN instance segmentation models. The study used aerial images collected from 2018 to 2019 covering rice planting-intensive areas of Changhua, Yunlin, Chiayi, and Tainan in central and southern Taiwan. The label file was created with four categories of rice growing, ripening, harvested stage, and other crops. The image pre-processing stage involved the modification of band information of the original aerial images using different vegetation indices such as NDVI, CMFI, DVI, RVI, and GRVI to create different image datasets. The study found that the ResNet-50 backbone performed better than the ResNet-101 in terms of mAP performance, with RGB + DVI, RGB + NIR, and RGB + GRVI image datasets performing the best with mAPs of 74.01%, 73.81%, and 73.72%, respectively. The models trained on RGB + CMFI images were recommended for the rice growing stage, RGB + NIR images for the rice ripening stage, and RGB + GRVI images for the rice harvested stage, with dice coefficients of 79.59%, 89.71%, and 87.94%, respectively. The detection and segmentation results can improve the efficiency and accuracy of rice production management by providing an understanding of rice growth status. This method can also be used for large-scale detection of any crops, reducing the burden on researchers and improving land use survey efficiency. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/87174 |
DOI: | 10.6342/NTU202210202 |
全文授權: | 同意授權(限校園內公開) |
電子全文公開日期: | 2028-02-13 |
顯示於系所單位: | 生物機電工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-1.pdf 目前未授權公開取用 | 16.31 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。