應用遙測技術於水稻田判釋之研究

Abstract

由於傳統以人工判釋航照資料，萃取水稻田資訊的方式過於耗費人力及時間，所以許多學者嘗試以衛星影像為材料，利用遙測技術來進行自動化判釋水稻田之研究。 執行自動化判釋水稻田面積時，大多是以遙測技術中的監督式分類方法來進行，但由於此法需以人工的方式來選取訓練樣區，多半無法完全達到客觀的要求，且因為 人因經驗及專業知識不足、或因地面資訊較複雜、或是肉眼不易辨識而選出偏頗之訓練樣區，進而影響判釋結果。因此，本研究的目的是設計出一自動化流程來選取 訓練樣區，利用遙測技術中的監督式分類方法進行分類，來提升判釋的準確度。 本研究室利用多時段SPOT衛星影像，來萃取雲林大埤鄉地區的水稻田耕作資訊。首先採用歷年農糧署所繪製水稻耕地坵塊圖，利用地理資訊系統中圖形套疊的技 術，將在研究區內連續九年（民國83至民國91年）皆為水稻及非水稻的坵塊位置擷取出來。然後將此萃取出來的歷年水稻與非水稻坵塊位置圖套疊2003年 SPOT衛星影像，擷取出水稻與非水稻像元，以非監督式分類方法得水稻與非水稻訓練樣區各20類。然後以此訓練樣區對一、二期水稻進行監督式分類，最後則 以農糧署提供之2003年水稻耕地坵塊圖進行全面性的正確性檢核。 研究顯示經由自動化選取訓練樣區方法可獲致良好判釋精度（一期稻作中整體精度為88％，kappa指標為0.75，二期稻作整體精度為92％，kappa 指標為0.85），優於以人工選取樣區之傳統監督性分類方法（一期稻作中整體精度為83％，kappa指標為0.68，二期稻作整體精度為 74％，kappa指標為0.63）。自動化選取訓練樣區方法，在有長時間且全面性的地真資料下，可改善人工選取訓練樣區之缺點，並提升判釋精度及效率， 達到以遙測影像自動化判釋水稻田之目的。

In the past years, remote sensing has been utilized to do the inventory of rice paddies by employ many technicians to manually interpret aerial photographs. This operation requires a large amount of photos and processing time. In this research, we propose an approach that integrates cultivating-field maps, multi-temporal SPOT images, and the spectral knowledge of rice growth to improve the efficiency of interpreting rice fields. In the traditional supervised classification procedures, the selection of the training sites is done by human-computer interaction. In addition, the selected training sites cannot involve all the spectral variations of rice fields usually. The traditional method is time consuming and the classification accuracy is usually not very good. Therefore, it is necessary to develop a procedure to automatic select training sites to cover all spectral variations of rice fields and further to improve the classification efficiency and accuracy. The results showed that the supervised classification accuracies using the automatic selection of training sites (first cultivating period: overall accuracy 88% and kappa 0.75, second cultivating period: overall accuracy 92% and kappa 0.85) are better then using the traditional human-computer interaction procedures (first cultivating period: overall accuracy 83% and kappa 0.68, second cultivating period: overall accuracy 74% and kappa 0.68). The bias induced by human while selecting training sites can be reduced by adopting automatic procedures.