透過空拍圖進行稻田區域影像分割與秧苗偵測之研究

A Study on Paddy Field Segmentation and Rice Seedling Detection from UAV Images

楊欣怡(Hsin-Yi Yang)

指導教授 ： 黃乾綱

國立臺灣大學/工學院/工程科學及海洋工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202001782

摘要

近年來，由於人口逐漸高齡化，台灣農業面臨勞動力嚴重短缺的問題。為了減輕農事生產作業量，進而降低勞動力需求，運用物聯網、感測元件、大數據分析等技術的智慧農業在台灣開始被大力推廣。其中，空拍機結合影像資料分析提供農民更便捷的作物成長監控方式，在農業上應用的需求日益增加。
本研究針對早期水稻秧苗之空拍圖，提出之方法其目的指在分割影像中的稻田區域與秧苗植株偵測，除此之外，提出方法也根據偵測之秧苗，標記其所屬稻田劃分區域，並完成相鄰影像與多張影像之間的秧苗對齊。研究主要分為四大部分進行實驗：（一）稻田區域影像分割、（二）秧苗偵測、（三）秧苗標記以及（四）秧苗對齊。第一部分使用三種影像分割演算法將輸入影像分割為不同的子區域，再利用分類資料集將每塊子區域分類為稻田或者非稻田兩類，最後合併稻田子區域得到稻田區域遮罩影像。第二部分利用不同植被指數產生前處理影像，於第一部分得到的稻田區域進行斑點偵測，所得之偵測結果即為秧苗。其中，此部分設計一個系統化的流程，使得每種前處理方法之偵測參數能被自動決定。第三部分藉由將輸入影像轉換至鑲嵌圖座標平面，依偵測秧苗轉換後的座標決定其所屬稻田區域並進行標記。最後第四部份提出一種基於秧苗的影像拼接方法來完成相鄰影像之間的秧苗對齊，再延伸到多張影像之間的秧苗對齊。
研究成果方面，第一部分使用的三種影像分割演算法，於最後的稻田區域分割結果，Ｆ度量值皆可達到95% 以上，其中以使用Felzenszwalb演算法之F度量值97.04% 為最佳。第二部分的秧苗偵測結果證明使用植被指數對於偵測結果有正面影響，而使用TGI具有最佳平均偵測結果，可達成平均F度量值93.84%，其中平均精確度為96.15%、平均召回率為91.96%。至於第三部分的標記秧苗所在稻田區域，其平均標記準確度可達99.2%。而第四部份在相鄰影像之間的秧苗對齊方面，使用提出方法所得之秧苗對數量，與原始方法相比平均可提升26.17%。在多張影像之間的秧苗對齊方面，ABC三張影像中，AC透過B間接達成影像拼接的方式，會比AC直接進行影像拼接所得結果更加精確。

關鍵字

精準農業 ； 空拍圖 ； 影像分割 ； 特徵偵測 ； 秧苗偵測 ； 影像拼接

並列摘要

Due to the aging of the population, Taiwan’s agricultural manpower has been in a large shortage in recent years. In order to reduce the labor demand of agricultural production, smart farming utilizing big data analysis, Internet of Things (IoTs), and sensors is of great interest. Particularly, combining with image data analysis, the inquiries of Unmanned Aerial Vehicles (UAVs) applied in agriculture is rising. It can provide farmers with better management on the process of growing crops.
The research proposed an approach to automatically label rice seedling from UAV images. There are 4 main parts in the proposed method: paddy field segmentation, rice seedling detection, rice seedling labeling, and rice seedling alignment. First, with the classification tool and three different segmentation algorithms: Watershed algorithm, Felzenszwalb’s algorithm and SLIC algorithm, paddy fields were extracted from the UAV images and the paddy mask image was computed. Second, by using blob detection with the paddy mask, rice seedlings were detected from different pre-processing images, which were generated by calculating vegetation indices (VIs) and image enhancement. Third, each detected rice seedling was labeled with an ID of paddy area by utilizing the homography matrix between the input UAV image and the mosaic. Last, the proposed method implemented an image stitching algorithm based on rice seedlings, and achieve rice seedling alignment.
For paddy field segmentation, all the three image segmentation algorithm can achieve more than 95% f-measure, where Felzenszwalb’s algorithm got the best performance with 96.26% precision, 97.88% recall and 97.04% f-measure. For rice seedling detection, most of the pre-processing images using VIs obtained better result than the pre-processing image which was just converted from RGB to grayscale. Among the pre-processing images, the one using TGI gained the best detection result with 96.15% precision, 91.96% recall and 93.84% f-measure. For rice seedling labeling, the result can achieve 99.20% accuracy. As for rice seedling alignment, the result shows that using seedling-based image stitching algorithm can increase 26.17% the numbers of rice seedling pairs between adjacent images. Besides, the result also shows that within multiple image stitching, the perspective transformation between the first image and the third image was more accurate by indirectly transforming with the second image.

並列關鍵字

precision agriculture ； UAV image ； image segmentation ； feature detection ； rice seedling detection ； image stitching

參考文獻

1. Wrigley, C., Batey, I., and Miskelly, D., Chapter 25 - Grain Quality: The Future is With the Consumer, the Scientist and the Technologist, in Cereal Grains (Second Edition), C. Wrigley, I. Batey, and D. Miskelly, Editors. 2017: Woodhead Publishing. p. 691-725.

2. Leonard, E.C., Precision Agriculture, in Encyclopedia of Food Grains (Second Edition), C. Wrigley, et al., Editors. 2016, Oxford: Academic Press. p. 162-167.

3. J. P. R, B., et al. Aerial mapping of rice crops using mosaicing techniques for vegetative index monitoring. in 2018 International Conference on Unmanned Aircraft Systems (ICUAS). 2018.

4. Ahmed, I., et al. Automatic Detection and Segmentation of Lentil Crop Breeding Plots From Multi-Spectral Images Captured by UAV-Mounted Camera. in 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). 2019.

5. Murata, K., et al. A Study on Growth Stage Classification of Paddy Rice by CNN using NDVI Images. in 2019 Cybersecurity and Cyberforensics Conference (CCC). 2019.