福爾摩沙衛星七號／氣象、離子層及氣候衛星第二星計畫（FORMOSAT-7/COSMIC-2）的無線電掩星（RO）觀測資料，在監測深層對流層以及為地球行星邊界層（PBL）提供關鍵見解方面具有顯著潛力。然而，在特定的熱力學條件下，自深層對流層檢索到的無線電掩星資料可能會產生嚴重的偏差。此偏差主要源於反演技術的限制、演算法的假設以及大氣因素的影響。本研究針對折射角偏差（BAB）與折射率偏差（REFB）的特性進行了檢驗，並基於這些特性，提出利用不同代用指標來評估特定區域偏差的統計技術。文中探討了兩種截然不同的方法：第一種是將傳統的迴歸模型與最小變異數估計法（MVE）進行整合；第二種則是應用創新的機器學習技術——類神經網路（NN）。研究結果顯示，採用基於最低可見高度（LSW）以及溫度與比濕關係的單一迴歸方法，成功捕捉到了與物理和熱力學條件相關的高振幅偏差。而用於結合各個獨立估計值的最小變異數估計法，則成為折射角偏差估計中最精準的方法。在類神經網路方面，本研究利用多層感知器（MLP）架構與反向傳播（BP）演算法，並以 ERA5 再分析資料進行訓練。結果表明，MLP-BP 架構能有效利用一組熱力學與觀測誤差變數來估計折射角偏差。整體而言，本研究凸顯了先進方法在改進無線電掩星偏差評估方面的潛力，能為更精準的大氣分析提供重要貢獻。;The potential of FORMOSAT-7/COSMIC-2 radio occultation (RO) measurements for monitoring the deep troposphere and offering crucial insights into the Earth′s planetary boundary layer (PBL) is significant. Yet， the RO data retrieved from the deep troposphere can have bias under specific thermodynamic conditions. The bias origins from retrieval technique limitation， algorithm assumption and atmospheric influences. Both characteristics of bending angle bias (BAB) and refractivity bias (REFB) are examined. Base on those characteristics， we propose statistical techniques that use different proxies to assess region-specific bias. Two distinct approaches are proposed: the integration of conventional regression-based models with the minimum variance method (MVE)， and the application of an innovative machine learning technique， the Neural Network (NN). The results shows that single regression approach employing LSW-based and temperature and specific humidity relationships， successfully captures high-amplitude bias associated both physical and thermodynamical conditions. The minimum variance estimation， which is used to combined individual estimations， emerges as the most precise method for bending angle bias estimation. The Neural networks (NNs) are trained on ERA5 reanalysis data using multi-layer perceptron (MLP) architecture and backpropagation (BP). The outcomes shows that MLP-BPs are adequate in estimating bending angle bias using a set of thermodynamic and measurement errors variables. This research highlights the promise of advanced methodologies in improving RO bias assessment for accurate atmospheric analysis.
