根據行政院農委會公布之統計資料顯示,農、林、漁及畜牧業在2006至2015年間受到颱風侵襲導致的災害損失共計約新台幣843億元。對應此份災損資料,本研究彙整颱風同期間全台風速與降雨數據,以超越機率閾值將數據分為5等級,並配合地區的土地利用概況,分析氣象變量、暴露量和災損間之關聯性。本研究亦利用淺層機器學習方法之倒傳遞類神經網路建立兩種空間尺度的災損預測模式。模式訓練結果顯示可有效模擬災損金額,Nash-Sutcliffe efficiency(NSE)最高為0.95,而全國平均可達0.7,而驗證結果顯示模式雖無法精確反映極端值,但在部分地區的NSE指數仍可達0.6以上,且合理地表現災損風險的空間分佈。本研究成果可予政府相關單位在颱風期間作為決策依據。
According to the statistics released by the Council of Agriculture, Executive Yuan, the aggregate typhoon losses in agriculture, forestry, fishery and animal husbandry were approximately $84.3 billion NTD from 2006 to 2015. This study acquired wind speed and rainfall data corresponding to each typhoon event in this period, and divided the data into five levels using exceedance probabilities as the thresholds. In conjunction with land-use data, correlations among weather variables, exposure, and losses were analyzed. Afterwards, a shallow machine learning method--back-propagation neural networks--was applied to the development of loss prediction models at different spatial scales. Calibration results showed that the models perform well at loss prediction with the highest Nash-Sutcliffe efficiency (NSE) as 0.95 and the national average ≥ 0.7. In validation, while upmost accuracy cannot be achieved for the prediction of extreme events, the cross-validated NSE can still reach 0.6 in certain regions, and a reasonable spatial distribution of typhoon risk is presented. This work is expected to provide guidance for relative government agencies prior to and during typhoon invasions.