以遙測影像的水稻田影像判釋:人工蜂群聚類演算法之研究

稻田在都市環境上對水土保持之影響主要有三: 區域性水田綠化環境、水田調洪功能與二氧化碳的吸收淨化空氣。在台灣水稻為台灣賴以維生的糧食作物，為了讓政府取得糧食政策制訂、產量推估與農民休耕或災後補助之依據來源，通常都會進行水稻種植面積的調查。而以往對於水稻田的判斷都是以實地探勘方式進行耕地坵塊圖（Rice Pattern）數化與編修。這些過程通常需要大量人力，因此透過遙測影像與合適的分類器足以改善現場探勘之困境。本研究以46筆水稻田的樣本與135筆的非水稻樣本，利用人工蜂群群集演算法建構判釋規則，研究採用多個植生指標與紋理資訊作為輔助因子，希望藉此提高判式的正確率，並建構一套有助於有效判釋水稻田的策略。本研究使用步驟可細分為(a)建立蜜蜂數量(b)計算蜜源適應度(c)計算各群心之離散度(d)進行迭代至終止條件，本研究採用隨機400點進行正確率之判釋，所得正確率結果為89.5%，在同樣設計的平行研究中，與蟻群聚類法進行比較後結果發現蜂群聚類演算法較佳，所以此方法對於水稻田的判釋有良好的成果。

關鍵字

蜂群聚類演算法；蟻群聚類演算法；影像分類

並列摘要

Paddy rice is the major crop of food in Taiwan. There are three main contributions on Taiwan: regional eco-friendly of environments, adjustment of floods and refresh the air pollution. The estimation of area is important since this information are related to the national food policy, yearly crop yields calculation and post-disaster reimburse. In the past, it is performed through field exploration and rice pattern revising, which is a large amount of human power and time-consuming works. A more economic manner to estimate paddy rice area is desired. Accordingly, this study aims to design a paddy rice classifier in which the area of paddy rice in a remote sensing image can be calculated. In this study, the ant-clustering algorithm is developed to paddy rice image classification. The advantage of this algorithm is to effective cluster data into groups for unsupervised learning. The advantage of this algorithm is to effective cluster data into groups. The study used 46 paddy rice samples and 135 non-paddy rice samples to construct the knowledge rules through Bee-based clustering. The study integrated texture information and vegetation indices as ancillary information to enhance the classification outcomes. We also build an effective strategy to identify the paddy rice. The steps are: (a) initialize the number of bee (b) compute the fitness of honey-source. (c) compute the divergence. (d) do iteration to final condition. We also use 400 point to count the accuracy rate. The result is 89.5%. In the parallel study, we also use ant-based clustering (ABC) to compare the outcome. It is found that BBC has the better classification performance than that of ABC.