• Journals

比較PS-InSAR與SBAS-InSAR技術應用於監測山區地表變位－以仁愛鄉為例

傳統崩塌監測技術與儀器受空間限制，本研究應用遙感之合成孔徑雷達干涉技術（Interferometric Synthetic Aperture Radar, InSAR）實現山區地表變位大範圍監測，採用永久散射體雷達干涉（Persistent Scatterer InSAR, PS-InSAR）與短基線子集差分干涉法（Small Baseline Subset-InSAR, SBAS-InSAR），以臺灣南投縣仁愛鄉為示範區，透過2017年間Sentinel-1A衛星共30幅升軌雷達影像，比較兩種多時序InSAR成果於地表變位監測的適用性，並與全球導航衛星系統（Global Navigation Satellite Systems, GNSS）數據進行相關性分析。結果顯示，PS-InSAR及SBAS-InSAR在LSAN測站皆呈現顯著正相關，相關係數分別為0.486及0.399，均方根誤差為5.004 mm及7.685 mm。SBAS-InSAR能有效反映山區之實際崩塌空間分布與地表變位情況，顯示該技術對山區崩塌監測更具優勢。

• Journals

UAV Path Determination for Visual Navigation using Deep Reinforcement Learning

Traditional path planning algorithms for Unmanned Aerial Vehicles (UAVs) primarily optimize for geometric metrics such as path length and energy efficiency. However, in GPS-denied environments, where external positioning is unreliable, the quality of visual localization is paramount for mission success. This study introduces a novel Deep Reinforcement Learning (DRL) framework designed to co-optimize the UAV path for both geometric efficiency and visual localization robustness. Specifically, our method integrates the density of matched image feature points, extracted from post-processed aerial imagery, directly into the planning process, ensuring the generated trajectory passes through visually rich areas that enhance navigation accuracy. To tackle the path planning challenge and address issues related to sparse rewards and unstable training, we employ an advanced DRL architecture: Noisy Dueling Double DQN with Prioritized Experience Replay (Noisy D3QN with PER). This integration leverages Double DQN to refine value estimation, Dueling DQN to improve generalization, PER to enhance sample efficiency, and Noisy Networks to promote robust and efficient exploration. The proposed framework is implemented within a simulated 2.5D environment with a customized reward function that considers both UAV state parameters and terrain features. Experimental results demonstrate that the method generates efficient, visually coherent, and dynamically smooth trajectories. Crucially, it enables path inference for multiple independent missions from various starting points after a single training session, achieving superior computational efficiency compared to traditional geometric planners. This highlights the potential of integrating visual features into a reinforcement learning-based UAV path planning to significantly enhance visual localization performance in complex environments.

• Journals

應用Geo-AI模型推估苗栗地區NO_2濃度及影響因子貢獻解析

本研究以苗栗縣為示範區，整合氣象、地面監測、衛星遙測、土地利用與通霄電廠排放等多源資料，建構逐時、50 m × 50 m解析度之NO_2濃度推估模型。採用XGBoost演算法SHAP遞增篩選機制，選取一小時滯後，模型R^2為0.80，RMSE為1.78 ppb。SHAP分析顯示NO_x與道路密度為最主要驅動因子，台電固定源亦具區域影響力。多層次驗證證實模型對時間、空間與高污染情境皆具穩健表現。研究展現Geo-AI於中小型縣市空品推估之應用潛力，亦為風險預警與污染治理提供量化依據。

• Journals

結合水文分析及UAV數值地表模型進行旱溪溢淹檢討

本研究主要貢獻為應用UAV數值地表模型提供精細的斷面資料，針對台中旱溪進行水文分析與洪水溢淹模擬。配合台中氣象站104年雨量資料，以五種機率分布模型推估一日最大暴雨量，經檢定與誤差分析，對數皮爾森三型為最佳模型。以符合旱溪實況之三角形單位歷線法進行洪水模擬，並利用HEC-RAS進行七種重現期距模擬。研究結果發現旱溪第四段於50年與100年洪水重現期距下有溢淹情形。本研究結合UAV數值地表模型提供精細的斷面資料，提升模擬精度與可靠度，分析成果可提供水利單位於河道整治與防洪規劃之參考。