稻米為台灣的主要經濟糧食作物，多種植於台灣的西部平原與東部花東縱谷。衛星影像為近年用於農作監測常用的資料，高時間解析度影像能捕捉稻作的生長情形，有利於稻作判釋，然而高時間解析度影像其空間解析度並不高，對於平均坵塊面積較小且破碎的台灣並不適用。目前衛星感測器難以同時具有兩種特性，因此需要結合不同的衛星進行影像融合，產製高時空時序融合影像。 本研究目的為因應台灣的小塊稻作坵塊，結合MODIS影像的8天時間解析度與Landsat影像的30公尺空間解析度，產製2012、2013年於研究區的多時序融合影像，並基於稻作物候特性判釋一期、二期稻作，以此檢核影像融合成果。主要流程分為五個步驟：第一步驟，對參與使用的衛星影像進行影像前處理，如幾何校正等。第二步驟，以STARFM (Spatial Temporal Adaptive Reflectance Fusion Model)融合改進方法對研究區MODIS與Landsat影像進行融合，產生高時空融合影像。第三步驟，以融合資料建構常態化植生指標 (Normalized Difference Vegetation Index, NDVI)的時序影像，並以小波轉換 (Wavelet Transform)濾除時序資料上的雜訊，保留稻作資訊。第四步驟，使用支援向量機 (Support Vector Machine, SVM)分類一期、二期稻作。第五步驟，以地真資料與統計資料對分類成果評估精度。 分類結果顯示與地真資料有高度一致性，整體而言，總體精度都在80%以上，Kappa值在0.65以上。與統計資料相比，R^2在0.85以上，RMSE (Root Mean Square Error)最多占統計面積的3%以內，分類成果多數的鄉鎮為誤判 (Commission)。因此以STARFM的融合改進方法產製多時序融合影像判釋稻作，約有80%的精度。 ;Rice is the main food crop in Taiwan, mostly grown in western plains and eastern Tai-wan. In recent years, satellite data are used for crop monitoring. High temporal resolution data can provide information of crop phenology. However, only high temporal resolution data are not sufficient for rice mapping in Taiwan, because rice fields here are generally small and fragmented. In this case, it is necessary to combine different satellite image, get-ting a high spatial resolution and high temporal resolution fusion image for rice mapping. The study aims to identify rice fields in areas in Taiwan using time-series MODIS (8-day)-Landsat (30m) fusion data in 2012 and 2013. The study consists of five steps: (1) Correct geometric and radiometric errors of Landsat data in data pre-processing; (2) Use the spatial temporal adaptive reflectance fusion model (STARFM) to blend the fusion data; (3) Construct smoothed normalized difference vegetation index (NDVI) time-series data by wavelet transform function; (4) Use support vector machine (SVM) to classify data; (5) Es-timate major rice crop area and assess mapping accuracies. The results indicate a high correlation between the mapping results and ground truth data. The overall accuracies are upper than 80%, and Kappa values are upper than 0.65. Compared with the government rice statics, the R2 are upper than 0.85, and the root mean square error (RMSE) are up to 3% of total rice statics. Most of the mapping results of town-ship level are commission. It means using time series fusion data to mapping rice is useful, it’s about 80% accuracy.
