永續計畫資料庫

Climate change and global geopolitical conflicts have jeopardized food security worldwide. Given Taiwan’s unique geopolitical situation and heavy reliance on food imports, along with the continuous shrink of farm land area, safeguarding national food production has become imperative. Among all, rice crop is the most critical one for maintaining people’s daily livelihood. Accurate prediction of annual rice yields is crucial for policy making and resource distribution, yet current existing models have not fully sufficed the accuracy level required. This proposal addresses this challenge by employing both optical and synthetic aperture radar (SAR) satellite imagery in a machine learning (ML) framework. There are two innovative elements in this proposal: multi-frequency SAR polarimetry (PolSAR) and canopy structure dynamics model (CSDM). To obtain multi-frequency SAR images suitable for polarimetric analysis, the proposal will attempt to task repeat-pass, high-resolution, full-polarimetric sensing modes over the target area in central western Taiwan. The optical and PolSAR information will then be linked to rice growth stages for feature selection, and then further to rice yields by using multiple ML models. In addition, a SAR-optical integrated CSDM will be created to fill the gaps when no observation data is available. The CSDM model will be examined separately for the use in yield prediction and utilized jointly within the ML models. If this project proves successful, we expect that the government will achieve better budget/resource allocation, import/export decision, crop price stabilization during ordinary years, water resource and pesticide management, as well as hazard mitigation plans, all critical to the long-term goals of sustainable agriculture.

氣候變化和全球地緣政治衝突危及了全球糧食安全。鑒於臺灣獨特的地緣政治局勢、對糧食進口的高度依賴以及農田面積的持續縮減，保障國家糧食生產已成為當務之急。其中，稻米作物對維持人⺠⽇常生活⾄關重要。準確預測年度稻米產量對於政策制定和資源配置⾄關重要，但現有模型尚未完全達到所需的準確性。本計畫透過在機器學習框架中結合光學和合成孔徑雷達衛星影像來應對這⼀挑戰。本提案中有兩個創新要素：多頻率合成孔徑雷達(SAR)極化分析應用和建構冠層結構動態模型。為了獲取適合極化分析的多頻率 SAR 影像，本計畫將嘗試進行衛星調度請求，以在臺灣中⻄部⽬標區域上空進行高解析度、全極化模式的重複觀測。我們擬將光學和 SAR 極化資訊與⽔稻生⻑階段聯繫起來進行特徵選擇，並使用多種機器學習模型進行稻米產量預測。此外，本計畫將試圖建立 SAR-光學集成的冠層結構動態模型，以填補無觀測資料時的空白。冠層結構動態模型將可單獨用於產量預測，亦可與機器學習模型共同使用。如果本計畫成功達成⽬標，我們預計政府將在預算與資源配置、進出口決策、正常年份的作物價格穩定、⽔資源與農藥管理、減災計畫方面等取得更好的成效，以進⼀步促成永續農業的⻑期⽬標。