水稻是臺灣地區的重要糧食作物,基於產量預估與災害補助依據所需,政府每年必須進行兩次水稻種植區域的調查。現行藉由航空像片搭配耕地坵塊圖,根據耕地坵塊所呈現之紋理以人工方式辨識該區塊是否為水稻田的做法相當耗費人力與時間。本研究嘗試結合耕地坵塊之地理資訊與多時段衛星遙測影像,利用各坵塊內之光譜植生指標隨時間的變化情形,以差分影像分類法(Differenced Image Classification)、時間剖面匹配法(Temporal Profile Matching)與波峰偵測法(Peak Detecting)進行自動化水稻田辨識作業。實驗成果顯示結合地理資訊可大幅提高分類之精度,而多時段衛星影像的運用,不但可再提昇分類精度,而且不需要訓練資料,可以達到自動辨識的目的。
Rise is the primary food crop in Taiwan. Rice inventory is therefore an important mission for the government to evaluate total rice production and provide a reference to subsidize economic loss caused by agricultural disasters. Currently, a photogrammetric surveying program manually interpreting aerial photos, are employed for the inventory work by the government. The procedures require a large amount of processing time and labor cost. In this research, a new approach to recognizing rice field automatically by integrating geographic information of land parcel and multi- spectral satellite images is proposed. Based on the fact that the vegetation coverage is changing during a rice season, the variation of the spectral properties and the resulting vegetation index provides the information to recognize rice fields. According to this idea, the methods proposed are Differeneced-Image-Classification (DIC), Temporal-Profile-Matching (TPM), and Peak-Detecting (PD). In the case of using a single image epoch in paddy classification, the use of geographic information will largely improve the accuracy. When multi-temporal images are used, it can recognize rice fields easily without the need of training data and can further improve the accuracy.