- 學位論文
類神經網路應用於颱風期間全台區域農業損失之研究
Investigation on the regional agricultural losses during typhoon periods by artificial neural networks
摘要
臺灣屬亞熱帶地區且位於西北太平洋颱風的主要侵襲路徑上，平均每年約有3~4次颱風影響到臺灣區域，近年因全球氣候變遷致使颱風侵台次數與強度有增加之趨勢，導致農民面對天氣驟變的災害損失與風險有明顯的增加；在農業損失方面又以稻米損失最為嚴重，且全台各縣市皆有稻田，故選用為研究標的。因此本論文希望能藉由農作物損失程度/金額探討並預測不同類型的颱風對臺灣的影響，進一步以稻米作物為標的，探究颱風造成全台區域性之損失；研究內容主要分為三部份，首先，分析歷年侵台的颱風特性，針對不同颱風類型進行相關性分析及探討各種災害損失在時間動態上的趨勢，反映出多元且複合的颱風災害特性。其次，根據上述分析結果，去除相關性較低的颱風特性資訊，作為類神經網路輸入項，並以自組特徵映射網路（Self-Organizing Feature Map, SOM）之聚類分析成果做為自組特徵輻狀基底函數（Self-Organizing Radial Basis Function Neural Networks, SORBFNN）網路之中心點，以建構颱風侵台時全台農作物損失與全台稻米損失之預報模式；模式結果顯示若以可能覆蓋臺灣面積取代七級風暴風半徑與是否登臺兩項颱風特性作為網路輸入項，則有最佳的預報成果。對於全台稻米損失預報，除了以可能覆蓋臺灣面積訊息作為輸入，適當加入侵台旬數與損失的隸屬度關係資訊確實能提昇模式推估的精確性。最後，因考量不同地區於稻米栽種的地理環境與社經條件上略有不同，使得颱風災害有明顯的空間差異。因此以Spearman相關性分析挑選出導致稻米損失的主要氣象因子，將臺灣分為三個區域並各自建立區域性的稻米損害預報模式。由模式成果中可發現，模式的輸入項對中南部地區稻米災害損失之影響遠超過北部和東南部。根據研究成果，希望能提升臺灣農作物於颱風時期之災害風險管理，並即時評估農業天然災害救助金的參考資訊。
並列摘要
Taiwan is located in subtropical zone and on the main typhoon track in the northwestern Pacific Ocean, and thus is with frequent typhoons events. On average, typhoons attack Taiwan more than three times per year because of global climate changes. In agriculture, farmers suffered from crop losses and increasing risks of damages due to climate changes. The prediction of rice losses is selected as the goal of this study because its coverage and losses is much more than other agricultural products’ in Taiwan. On the other hand, the impacts of different types of typhoons on Taiwan are also investigated according to the amount of agricultural losses. The content of this study was divided into three steps; first, the characteristic of historical typhoons and the correlation between different types of typhoons and various losses were analyzed. Second, variables highly related to agricultural losses were identified as inputs to artificial neural networks according to aforementioned results. The Self-Organizing feature Map (SOM) network was then applied to classifications which were adopted as the central points of Radial Basis Function Neural Networks (RBFNN). The SORBFNN was built for separately predicting the agricultural losses and rice losses during typhoon periods. The results indicate that the best performance could be obtained when the radius of 7th Beaufort wind scale and typhoon landfall status were replaced by possible coverage. As far as the prediction of rice losses is concerned, the accuracy of model outputs can be increased by adding the membership functions of time and losses besides possible coverage. Finally, the spatial distribution of agricultural losses is significantly distinct because rice was planted in various counties with different environment and social financial conditions. Therefore, Taiwan was classified into three sub-regions and the major climate factors resulting in rice losses in each sub-region were also selected by using Spearman correlation analysis. Results obtained from three regional rice-loss models show that the selected inputs have higher impact on the model built in central region. According to the results achieved, the model is expected to increase the risk management of agricultural losses during typhoon periods and to immediately provide the information on the agricultural subvention when natural disaster occurs.