Please use this identifier to cite or link to this item:
|Title: ||應用Sentinel-1A合成孔徑雷達影像於臺灣中部水稻田之判釋;Rice crop classification using Sentinel-1A SAR data in Central Taiwan|
|Keywords: ||水稻;合成孔徑雷達影像;Sentinel-1A;NDSI;rice;Synthetic Aperture Radar image;Sentinel-1A;NDSI|
|Issue Date: ||2016-10-13 14:45:29 (UTC+8)|
|Abstract: ||水稻是臺灣重要的糧食作物之一，也是大多數農村人口主要的生計來源。臺灣水稻種植的面積佔臺灣總面積的5%，約166,616公頃，所以水稻監測是必需且勢在必行的工作。近年來，光學衛星影像，如SPOT, MODIS及FORMOSAT-2，已經被廣泛地應用於臺灣的水稻監測，但是臺灣水稻的生長季節與雨季相符，雲霧覆蓋成為利用光學衛星影像判釋水稻的絆腳石；合成孔徑雷達屬於主動式的微波感測器，可以穿透雲霧且不受天候的影響進行監測，如RADARSAT-2與ERS-2，而此類型的影像通常較光學衛星影像昂貴。2014年Sentinel-1A升空後，提供全球免費的Sentinel-1A合成孔徑雷達VV與VH極化影像。利用多時序合成孔徑雷達影像的後向散射係數計算Normalized Difference Sigma-naught Index (NDSI)指標，可以了解地物在不同時期的變化。本研究利用多時序Sentinel-1A合成孔徑雷達VV與VH極化影像，配合NDSI發展一套自動化判釋的方法，分析臺灣中部水稻種植之範圍，並且探討不同極化與NDSI於判釋水稻的差異性。主要分為五個步驟：第一步驟，將原始影像進行前處理，包括輻射校正、幾何校正及雜訊濾除。第二步驟，根據水稻物候計算NDSI並定義靜態NDSI與動態NDSI。第三步驟，因為中部地區水稻種植區域較為集中，利用此空間特性進行多尺度影像分割。第四步驟，利用期望值最大化法與NDSI閾值發展一套自動化方法，判釋水稻與非水稻的種植區域。第五個步驟，利用政府水稻地真資料與水稻種植面積統計資料進行判釋成果的精度評估。研究成果顯示，升軌模式靜態NDSI VH與VV分類結果的總體精度分別為85.1%和65.1%，Kappa係數分別為0.69和0.29，動態NDSI 分類結果的總體精度分別為92.1%和78.2%，Kappa係數分別為0.85和0.56。降軌模式靜態NDSI VH與VV分類結果的總體精度分別為88.1%和69.1%，Kappa係數分別為0.75和0.37，動態NDSI 分類結果的總體精度分別為92.0%和81.1%，Kappa係數分別為0.84和0.62。比較分類結果面積與政府統計資料，升軌與降軌模式VH極化的分類結果的RMSE僅占總面積1%以下，VV極化的RMSE則相對較高。顯示不論是升軌模式或降軌模式，利用動態NDSI與VH極化可以獲得較佳的水稻判釋成果，且本研究發展的方法能有效地應用於水稻判釋之相關研究。;Rice is the most important staple food crop and the primary source of livelihoods for the majority of rural populations in Taiwan. The area allocated for rice cultivation accounts for approximately 5% (166,616 ha) of the total cultivating area. Therefore, rice monitoring is a crucial activity in Taiwan due to official initiatives. In recent years, optical satellite data acquired from sensors such as SPOT, MODIS, and FORMOSAT-2 satellites have been widely used for rice crop classification. However, because rice is mainly cultivated in the rainy season in Taiwan, the optical satellite data reveal challenges due to cloud cover during this season. The synthetic aperture radar (SAR) such as Radarsat-2 and ERS-2 data, which can penetrate clouds and operate in all weather conditions, are generally expensive. With the launch of Sentinel-1A in 2014, it is possible to acquire free VV and VH polarization data in the study region. This study aims to develop a mapping approach to delineate rice cultivation areas in Central Taiwan using the Normalized Difference Sigma-naught Index (NDSI) calculated from the time-series Sentinel-1A VV and VH polarization data. Two types of NDSI were used in this study: (1) static NDSI, which was calculated using only two images of sowing and heading dates, and (2) dynamic NDSI, which was calculated using the time series of images. An assessment of the applicability of VV and VH polarization data of Sentinel-1A and different types of NDSI for rice crop mapping was also performed. The methodology of this study comprises five steps: (1) data pre-processing, including radiometric, geometric corrections, and speckle noise filtering of the backscattering coefficient of VV and VH polarization data, (2) calculation of static and dynamic NDSI, (3) image segmentation, (4) threshold-based rice classification using the expectation-maximization algorithm, and (5) accuracy assessment of the mapping results using the ground rice reference data and government rice area statistics. The mapping results achieved from the ascending static NDSI VH and VV polarization data indicated the overall accuracies of 85.1% and 65.1% and Kappa coefficients of 0.69 and 0.29, respectively, while those from the ascending dynamic NDSI VH and VV polarization data indicated the better overall accuracies of 92.1% and 78.2% and Kappa coefficients of 0.85 and 0.56, respectively. Similarly, the results obtained the descending static NDSI VH and VV polarization data (overall accuracies of 92.0% and 81.1% and Kappa coefficients of 0.84 and 0.62) were better than those from the ascending static NDSI VH and VV polarization (overall accuracies of 88.1% and 69.1% and Kappa coefficients of 0.75 and 0.37). Furthermore, the RMSE value obtained by comparing the ascending and descending VH classification results and government statistics was lower than 1%, in all cases. Compared to VH polarization, the RMSE of VV polarization results was higher than that of VH results. This study demonstrates that the potential applicability of VH polarization data using the dynamic method for rice crop mapping in the study region. The methods were thus proposed for rice monitoring in the study region.|
|Appears in Collections:||[遙測科技碩士學位學程] 博碩士論文|
Files in This Item:
All items in NCUIR are protected by copyright, with all rights reserved.