ENGLISH

29卷/3期

29卷/3期

華藝線上圖書館

Pages:

129-149

論文名稱

三維建物線框模型之無人機影像自動重建

Title

Automatic UAV Image Reconstruction for 3D Building Wireframe Models

作者

黃郁翎、莊子毅

Author

Yu-Ling Huang, Tzu-Yi Chuang

中文摘要

三維建物圖資在智慧城市規劃、管理和能源評估中扮演著重要角色。然而,由於作業繁瑣且自動化不足,針對既有建物構建精確的三維模型依然充滿挑戰。本研究提出基於多視角無人機影像的演算策略,生成具備側向幾何細節的三維線框模型,可做為進行既存建物三維房屋模型建置之基礎,提升作業效率並降低成本。演算程序運用預訓練的角點檢測模型及提出的角點萃取演算,採用「由粗到細」的策略實現角點定位。同時,運用虛擬角點重建策略來降低都市UAV影像中無可避免的遮蔽與數據缺失影響。實驗結果顯示,演算策略可適應具曲線形之建築結構,建築角點平均精度約為30 cm,並可達到98%的線框重建完整度。

Abstract

3D building data is vital in smart city planning, management, and energy assessment. However, constructing accurate 3D models for existing buildings remains challenging due to the labor-intensive processes and insufficient automation. This study proposes an algorithmic strategy based on multi-view UAV imagery to generate 3D wireframe models with detailed lateral geometric features, serving as a foundation for constructing 3D building models of existing structures. This approach aims to improve operational efficiency and reduce costs. The algorithm employs a pre-trained corner detection model and a novel corner extraction algorithm, utilizing a "coarse-to-fine" strategy to achieve precise corner localization. Additionally, a virtual corner reconstruction strategy is employed to mitigate the inevitable occlusion and data loss in urban UAV imagery. Experimental results demonstrate that this algorithmic strategy adapts well to buildings with curved architectural structures, achieving an average corner localization accuracy of approximately 30 cm and up to 98% completeness in wireframe reconstruction.

關鍵字

多視角無人機影像、影像建模、自動化建物線框重建、線框模型、深度學習

Keywords

Multi-View UAV Imagery, Image Modeling, Automated Building Wireframe Reconstruction, Wireframe Models, Deep Learning

附件檔名

華芸線上圖書館

https://www.airitilibrary.com/Article/Detail/10218661-N202410020017-00001/

備註說明

N / A

Pages:

151-164

論文名稱

結合溫室氣體排放、綠蔽度衛星影像與土地利用資料的環境溫度機器學習預測模型開發

Title

Development of an Ambient Temperature Prediction Model Using Machine Learning by Integrating Greenhouse Gas Emissions, Vegetation Index Satellite Images, and Land Use Data

作者

張皓庭、陳映融、柳婉郁、吳治達

Author

Hao-Ting Chang, Yinq-Rong Chern, Wan-Yu Liu, Chih-Da Wu

中文摘要

本研究綜合考量了溫室氣體、環境和人為活動相關變數,以利用大數據與五種機器學習演算,包含:隨機森林 (RF)、梯度提升 (GBR)、輕量梯度提升 (LGBMR)、類別提升 (CBR) 和極限梯度提升 (XGBoost)來建立兩種溫室氣體CO2和CH4推估環境溫度的模型,其中LGBMR模型在CO2方面表現最佳,而XGBR模型在CH4方面效果較好。CO2和CH4推估環境溫度模型的表現,R2值分別為0.993和0.999。SHAP值的分析確認了溫室氣體濃度、濕度、風速等因素對預測的關鍵影響。本研究成果為溫室氣體減排策略提供了新的評估方法,並為全球氣候變化研究與政策制定提供了重要參考,凸顯了跨學科合作的重要性。

Abstract

This study integrated greenhouse gases, environmental, and anthropogenic variables, utilizing big data and five machine learning algorithms, including Random Forest (RF), Gradient Boosting (GBR), Light Gradient Boosting Machine Regressor (LGBMR), CatBoost Regressor (CBR), and eXtreme Gradient Boosting (XGBoost), to establish models for estimating ambient temperatures based on two greenhouse gases, CO2 and CH4. The LGBMR model performed best for CO2, while the XGBR model showed better performance for CH4. The R2 values for the CO2 and CH4 estimation models were 0.993 and 0.999, respectively. Analysis of SHAP values confirmed the significant influence of greenhouse gas concentration, humidity, wind speed, and other factors on predictions. The findings of this study offer new evaluation methods for greenhouse gas emission reduction strategies and provide crucial insights for global climate change research and policy-making, highlighting the importance of interdisciplinary collaboration.

關鍵字

環境溫度、二氧化碳、甲烷、機器學習預測模型、機器學習演算法

Keywords

Ambient Temperature, Carbon Dioxide, Methane, Machine Learning Predictive Model, Machine Learning Algorithms

附件檔名

華芸線上圖書館

https://www.airitilibrary.com/Article/Detail/10218661-N202410020017-00002/

備註說明

N / A

Pages:

165-176

論文名稱

基於土地利用迴歸之機器學型模型分析新冠肺炎三級警戒政策對臺灣工業城市細懸浮微粒之影響

Title

Impacts of the Level 3 Alert Brought by COVID-19 on Fine Particulate Matter of an Industrial City in Taiwan Using a Land-Use Based Machine Learning Model

作者

蘇均珺、翁佩詒、曾于庭、李佳禾、吳治達

Author

Jun-Jun Su, Pei-Yi Wong, Yu-Ting Zeng, Chia-Ho Lee, Chih-Da Wu

中文摘要

COVID-19疫情對全球帶來巨大衝擊,臺灣政府於2021年5月19日宣布三級警戒,限制民眾活動。本研究旨在評估警戒期間臺灣工業城市PM2.5濃度之變化。本研究以代表性的工業城市高雄市為研究區域,蒐集1994至2020年的空污觀測數據和地理變量,利用土地利用迴歸和逐步變量選擇建立模型、選取重要變數,再使用不同機器學習演算法建立模型,其中結果以Random Forest (RF) 演算法的模型表現最佳,R2達0.95;推估成果顯示封鎖期間空氣品質改善,高雄市平均PM2.5濃度為18.1μg/m3,低於警戒前19.9 μg/m3。Paired t-test結果顯示差異達到統計顯著水準 (p值<0.001),各土地利用區域 (居住區、工業區、街道和綠地) 亦呈現一致結果。

Abstract

The COVID-19 epidemic has brought significant changes to human activities worldwide, including in Taiwan. On May 19, 2021, the government announced a level 3 alert to restrict public movement. This study aims to assess the impact of the lockdown policy on PM2.5 concentrations in Taiwan's industrial city, Kaohsiung. Daily PM2.5 observations and geographic data from 1994 to 2020 were collected. A land-use regression model, combined with stepwise variable selection, was used to identify important factors affecting PM2.5 variability. These predictors were used to develop machine learning models with algorithms such as Random Forest (RF), which showed the best performance with an R2 of 0.95. Paired t-tests indicated that PM2.5 levels were significantly lower during the alert (18.1 μg/m3) compared to before (19.9 μg/m3), with consistent results across residential, industrial, street, and green areas (p < 0.001).

關鍵字

空氣污染、細懸浮微粒物 (PM2.5)、基於土地利用的機器學習模型、COVID-19、三級警戒

Keywords

Air Pollution, Fine Particulate Matter (PM2.5), Land-Use Based Machine Learning Model, COVID-19, Level 3 Alert

附件檔名

華芸線上圖書館

https://www.airitilibrary.com/Article/Detail/10218661-N202410020017-00003/

備註說明

N / A

Pages:

177-189

論文名稱

運用生成對抗網路產製超解析衛星影像之先期評估

Title

The Preliminary Evaluation of Generating Super-Resolution Satellite Images Using Generative Adversarial Networks

作者

張庭韶、蔡富安

Author

Ting-Shao Chang, Fuan Tsai

中文摘要

本研究探討使用生成對抗網路 (GAN) 模型提升衛星影像的空間解析度,以解決因拍攝角度、天氣狀況及感測器限制導致的解析度下降問題。研究使用中、高及超高解析度影像進行降解析處理,並透過 GAN 進行訓練生成超解析影像。GAN 訓練過程中,生成器負責將低解析度影像重建為高解析度影像,判別器則區分生成影像與真實影像的差異。本研究亦使用 VGG-19 預訓練模型進行特徵提取,提升生成影像品質。實驗結果顯示,隨訓練次數增加,影像細節變得更為清晰及銳利化,且在結構相似性指標上優於傳統方法。然而,影像生成過程中出現色彩偏移及偽影現象。為改善此問題,建議進行更深層次訓練或使用後處理技術,並優化模型架構,如移除 Batch Normalization。綜上所述,GAN 模型具有提升衛星影像解析度的潛力,惟仍尚有影像色彩偏移及?影問題,未來可針對模型穩定性和影像後處理從事進一步改進與優化。

Abstract

This study explores the use of Generative Adversarial Networks (GAN) to enhance the spatial resolution of satellite images, addressing resolution degradation caused by factors such as off-nadir angles, weather conditions, and sensor limitations. The research utilizes medium, high, and very-high-resolution images, applying downsampling and training the GAN to generate super-resolution images. During GAN training, the generator reconstructs low-resolution images into super-resolution ones, while the discriminator distinguishes between generated and real images. The study also utilizes the VGG-19 pre-trained model for feature extraction to improve image quality. Experimental results show that image details become sharper as training progresses, and the GAN outperforms traditional methods in terms of structural similarity. However, issues like color shifts and artifacts emerged during image generation. To address these problems, the study recommends deeper training, post-processing techniques, and model optimizations such as removing Batch Normalization. Overall, while GAN models show potential for enhancing satellite image resolution, further improvements are needed to resolve color shifts and artifacts, focusing on model stability and post-processing.

關鍵字

超解析、衛星影像、生成對抗網路、深度學習

Keywords

Super Resolution, Satellite Imagery, Generative Adversarial Network, Deep Learning

附件檔名

華芸線上圖書館

https://www.airitilibrary.com/Article/Detail/10218661-N202410020017-00004/

備註說明

N / A

更多活動學刊