DSpace community: 環境科技博士學位學程

Monitoring Coal Mine Reclamation Compliance Using Deep Learning Analysis on Multitemporal Satellite Imagery

title: Monitoring Coal Mine Reclamation Compliance Using Deep Learning Analysis on Multitemporal Satellite Imagery abstract: 監測煤礦開採與復育活動對於確保環境問責與推動永續資源管理至關重要。遙感技術提供了一種強大的方法，能夠在無需大量實地調查的情況下觀測大範圍的地表變化。本論文提出了一個多時期深度學習框架，利用衛星影像系統性地監測並評估印尼南加里曼丹地區煤礦復育的合規情況。該研究分為兩個階段，反映了從初步研究到後續研究的方法進展。在第一階段，整合了 Sentinel-2 多光譜影像，使用 U-Net 分割模型來分類礦區與非礦區。從礦區轉變為非礦區的區域被解釋為復育地。雖然此方法能有效偵測地表擾動與恢復的整體模式，但將所有裸地都歸類為礦區，限制了其區分不同復育階段或具體地表狀況的能力。為了解決這一限制，第二階段通過僅使用 Sentinel-2 影像，結合多種光譜指數和綜合公式來改進該框架。利用基於 U-Net 的深度學習模型，將五種地表組成進行分類，包括表土層、次表土層、植被、煤層和水體，整體分類準確率達到 0.94，Kappa係數為 0.91。這一細緻的分類使得礦區與復育過程的追蹤更加精確，能夠識別次表土層暴露為開採活動，並將次表土層或煤層轉變為植被或表土層視為復育進展。從 2016 年到 2021 年的時序分析顯示，2019 年礦區大幅擴展，隨後在 2020 年復育活動顯著增加。將分類結果與煤礦許可邊界進行整合後，計算出合規比率（CR）介於 0.32至 1.44之間，反映了九個礦區許可持有者之間的差異。同時，還建立了復育活動指數（RAI），作為一種簡單的量化比較方法，用以檢驗其年度趨勢是否與深度學習導出的地表變化一致，結果顯示高度相關。總體而言，本研究提出的多時期深度學習框架證明了將衛星遙測與空間分析方法相結合，在礦區與復育監測方面的準確性、可擴展性與透明性，為基於資料的環境治理提供了有力支持。;Monitoring coal mining and reclamation activities is essential for ensuring environmental accountability and sustainable resource management. Remote sensing provides a powerful means of observing large scale land surface changes without requiring extensive field surveys. This dissertation develops a multitemporal deep learning framework using satellite imagery to systematically monitor and assess reclamation compliance in coal mining regions of South Kalimantan, Indonesia. The research is conducted in two stages, reflecting methodological advancements from the initial to the subsequent study. In the first stage, Sentinel-2 multispectral imagery were integrated to classify mining and non-mining areas with U-Net segmentation. Changes from mining to non-mining areas were interpreted as reclamation. While this approach effectively detected general patterns of surface disturbance and recovery, it generalized all barren land as mining, limiting its ability to distinguish detailed reclamation stages or specific surface conditions. To address this limitation, the second stage refined the framework by employing only Sentinel-2 imagery with multiple spectral indices and composite formulations. A U-Net based deep learning model was trained to classify five surface components topsoil, subsoil, vegetation, coal bodies, and water bodies with an overall accuracy of 0.94 and a Kappa coefficient of 0.91. This detailed classification enabled more precise tracking of mining and reclamation processes, identifying subsoil exposure as mining activity and transitions from subsoil or coal bodies to vegetation or topsoil as reclamation progress. Temporal analysis from 2016 to 2021 revealed substantial mining expansion in 2019, followed by a sharp increase in reclamation activity in 2020. The integration of classification results with coal mining permit boundaries produced CRs (CR) ranging from 0.32 to 1.44, reflecting variations among nine permit holders. A RAI was also developed as a simple quantitative comparison method to examine whether its annual trends align with the deep learning derived spatial changes, showing strong correspondence. Overall, the proposed multitemporal framework demonstrates the effectiveness of combining satellite based deep learning and spatial analytics for accurate, scalable, and transparent monitoring of mining and reclamation dynamics, supporting data-driven environmental governance.

Integrating Remote Sensing Data for Drought Assessment in Taiwan

title: Integrating Remote Sensing Data for Drought Assessment in Taiwan abstract: 春季乾旱對台灣造成重大影響，對當地的農業、工業和民生領域均有影響。這些乾旱尤其令人擔憂，因為它們對本地和全球經濟都有深遠影響，特別是半導體產業，這是全球科技市場的基石。這一點在2021年春季，台灣積體電路製造股份有限公司（TSMC）的晶片生產中斷中尤為明顯。儘管這些乾旱的嚴重性，當地研究在整合各種乾旱指數方面仍存在明顯差距，這對於全面的乾旱風險評估和管理非常重要。本論文針對這一差距，通過評估春季乾旱的特徵並提出一個創新框架，利用集成學習方法全面評估台灣的乾旱風險。首先，第3章採用了標準化乾旱指數（SDI）、標準化溫度指數（STI）和歸一化差異水指數（NDWI），研究了1982年至2021年間台灣春季乾旱的特徵和潛在氣候變化影響。該章節結合了遙測技術、統計分析和機器學習技術，並通過地面驗證的遙測數據進行強化。研究結果顯示，台灣的乾旱動態與全球典型模式不同。台灣的水文乾旱對降雨赤字的反應迅速，相關性強（r > 0.8），而農業乾旱的相關性適中（0.4 < r < 0.6）。此外，儘管農業乾旱有所減少，但自2000年代初以來，氣象和水文乾旱有所增加，並出現顯著的變化點；台灣中部和南部地區成為乾旱熱點，每4至6年經歷顯著的溫度-乾旱一致性循環。本章強調了改進乾旱管理策略的重要性，這對台灣的工業經濟尤為重要。其次，第4章提出了一個綜合框架來評估台灣的乾旱風險，結合了分析網絡過程（ANP）和人工神經網絡（ANN）。本研究的目標是繪製詳細的乾旱風險地圖，同時考慮乾旱災害對農業、工業和民生領域的社會經濟影響。最終的乾旱風險地圖通過實地調查和統計分析進行全面驗證，在評估作物損害、面積轉換和預估經濟損失方面，驗證準確率達到0.717到0.851之間。本章展示了ANP-ANN集成方法的有效性，證明了其在各種生態和社會經濟條件下迅速且準確預測乾旱風險的穩健性。最後，第5章總結了本論文的主要發現和貢獻，並討論所用方法的擴展潛力，以應對該領域的新挑戰。 ;Spring droughts in Taiwan significantly impact the socio-economic, affecting local agriculture, industry, and domestic sectors. These droughts are particularly concerning due to their far-reaching implications for both local and global economies, especially the semiconductor industry, a cornerstone of global technology markets. This is particularly evident in the disruption of chip production by Taiwan Semiconductor Manufacturing Company Limited (TSMC) in Spring 2021. Despite the critical nature of these droughts, there is a notable gap in the integration of various drought indices in local studies, which is essential for comprehensive drought risk assessment and management. This dissertation addresses this gap by assessing the characteristics of spring droughts and proposing an innovative framework that utilizes an ensemble learning method to comprehensively evaluate drought risk in Taiwan. Firstly, Chapter 3 employs Standardized Drought Indices (SDI), the Standardized Temperature Index (STI), and the Normalized Difference Water Index (NDWI) to investigate the characteristics of spring droughts from 1982 to 2021 in Taiwan. It utilizes a combination of remote sensing, machine learning techniques, and statistical analysis, enhanced by the assimilation of ground-validated remote sensing data. The results reveal that drought dynamics in Taiwan are distinct from typical global patterns. Hydrological droughts in Taiwan show a rapid response to rainfall deficits, with strong correlations (r > 0.8), whereas agricultural droughts demonstrate moderate correlations (0.4 < r < 0.6). Additionally, while agricultural droughts have seen a decline, meteorological and hydrological droughts have increased since the early 2000s, featuring a significant change point. The central and southern regions of Taiwan emerge as drought hotspots, experiencing notable air temperature-drought coherence in 4–6-year cycles. This chapter underscores the importance of advancing drought management strategies, particularly vital for Taiwan′s industrial economy. Secondly, Chapter 4 introduces a comprehensive framework to assess drought risk in Taiwan, utilizing a combination of the Analytic Network Process (ANP) and Artificial Neural Network (ANN). The research objectives of this study are to generate a detailed map of drought risk while considering the impacts of drought hazards on the socio-economic domains of agriculture, industry, and domestic sectors. The final drought risk map was thoroughly validated through fieldwork and statistical analysis, achieving high validation accuracies ranging from 0.717 to 0.851 in evaluating crop damage, area conversion, and estimated economic losses. This chapter highlights the effectiveness of the ANP-ANN ensemble method, proving its robustness in rapidly and accurately predicting drought risk under various ecological and socioeconomic conditions. Finally, Chapter 5 concludes the dissertation by summarizing its principal discoveries and contributions. The chapter also discusses the potential for expanding the methodologies employed, to address emerging challenges in the field.

Advancements in Satellite-Based Drought Monitoring Methods: Novel Indices and Their Applications at Various Scales

title: Advancements in Satellite-Based Drought Monitoring Methods: Novel Indices and Their Applications at Various Scales abstract: 乾旱是一種嚴重的自然災害，對全球的生態系統、環境和人類生活產生廣泛的影響。近期台灣的乾旱情況引起了重大關注，尤其是對半導體晶片生產這樣的重要產業。本研究使用衛星基礎指數來監測及分析乾旱狀態，因為準確乾旱評估在有效的水資源管理中起著關鍵作用。溫度-植被乾旱指數（TVDI），使用經驗技術結合地表溫度（LST）和植被覆蓋分數（FVC），被廣泛使用。然而，其在植被稀疏地區的適用性受到限制，促使我們探索替代方法。最近發展的溫度-土壤水分乾旱指數（TMDI），用標準化差異潛熱指數（NDLI）替代植被指數，顯示出做為一種合適替代品的潛力。研究的第一個關鍵部分介紹了使用從NDLI衍生出的新的表面水分可用性（FSWA）對TMDI的改進。該部分深入探討了在LST-FSWA空間內精煉邊緣選擇以觀察乾旱。已經採用了一種實用的方法來準確地識別這個空間內的乾和濕邊緣。為了評估TMDI的可靠性，該研究使用各種指標進行了全面的評估，包括由地表能量平衡算法（SEBAL）衍生的蒸散量（ET）、作物水分壓力指數（CWSI）、初級生產力（GPP）和降水數據。結果顯示，TMDI與SEBAL衍生的CWSI、ET和GPP之間的相關性高於TVDI。此外，TMDI與降水之間的強關聯突顯了其捕捉乾旱模式的有效性。該部分還提出了一個標準的TMDI閾值，用於評估2014年至2021年台灣西南部的乾旱。本文的後續部分應用新穎的 FSWA 來監測全國範圍內的農業缺水和乾旱狀況。 FSWA 與其他兩個指數一起用於分析 2001 年至 2022 年澳洲的年度乾旱模式。在澳洲的大多數農業地區，FSWA 顯示出與土壤濕度 (SM)、ET 和降雨量的強大時間相關性。鑑於SWAT的簡化計算和全國範圍的適用性，其實際利用已在全球範圍內進行評估。研究發現SWAT可以代表土壤水分並生成高分辨率的乾旱地圖。此外，SWAT被應用於評估2011年至2022年的全球乾旱狀況。基於SWAT指數的乾旱分佈和趨勢分析顯示出不同土地覆蓋類型之間的廣泛變化。總的來說，TMDI和SWAT都作為實際運行的乾旱指數。先進的TMDI，依賴於僅兩個變量的整合：FSWA和LST，在區域範圍的乾旱監測中表現出色，特別是在農業區。另一方面，由於其更廣泛的規模適用性和更簡單的方法，基於衛星的SWAT指數提供了一種實用的替代方案。;Drought, a detrimental natural disaster, has extensive impacts on ecosystems, the environment, and human life globally. Its recent emergence in Taiwan has sparked significant concerns, especially for vital industries like semiconductor chip production. This dissertation uses satellite-based indices to monitor drought status, given the crucial role of accurate drought assessment in effective water management. The Temperature-Vegetation Dryness Index (TVDI), which combines Land Surface Temperature (LST) and Fractional Vegetation Cover (FVC) using empirical techniques, is widely utilized. However, its applicability is limited in regions with sparse vegetation, prompting the exploration of alternative methods. The recently introduced Temperature-Soil Moisture Dryness Index (TMDI), which substitutes the vegetation index with the Normalized Difference Latent Heat Index (NDLI), shows potential as a suitable alternative. The first key section of the dissertation presents advancements in the TMDI using the new Fractional Surface Water Availability (FSWA) derived from the NDLI. This section delves into refining edge selection within the LST–FSWA space to observe drought. A practical method has been adopted to accurately identify the dry and wet edges within this space. To evaluate the reliability of TMDI, the study conducted comprehensive assessments using various indicators, including the evapotranspiration (ET), Crop Water Stress Index (CWSI) derived from the Surface Energy Balance Algorithm for Land (SEBAL), Gross Primary Productivity (GPP), and precipitation data. The results reveal high correlations between the TMDI and SEBAL-derived CWSI, ET, and GPP, surpassing those obtained with TVDI. Moreover, strong associations between the TMDI and precipitation underscore its effectiveness in capturing drought patterns. This section also proposes a standard TMDI threshold for assessing drought in southwestern Taiwan from 2014 to 2021. The subsequent section of this dissertation applies the novel FSWA to monitor agricultural water stress and drought status at the national scale. The FSWA is employed alongside two other indices to analyze annual drought patterns in Australia from 2001 to 2022. The analyses reveal high correlations between the FSWA against soil moisture (SM), ET, and rainfall. Across most agricultural regions in Australia, the FSWA shows robust temporal correlations with the SM, ET, and rainfall. Furthermore, given the SWAT’s simplified calculations and wide applicability, its practical utilization has been evaluated at the global scale. It is found that the SWAT can represent SM and generate high-resolution drought maps. The SWAT was applied to assess global drought conditions from 2011 to 2022. The analysis of drought distributions and trends based on the SWAT index exhibited wide variations across different land cover types. In conclusion, both the TMDI and SWAT function as practically operational drought indices. The advanced TMDI, relying on the integration of only two variables: FSWA and LST, excels in regional-scale drought monitoring, particularly in agricultural zones. On the other hand, the satellite-based SWAT index, due to its broader scale applicability and more straightforward methodology, offers a viable alternative to empirical models.

利用遙感和機器學習技術量化城市熱島現象;Quantification of urban heat island phenomenon with remote sensing and machine learning techniques

title: 利用遙感和機器學習技術量化城市熱島現象;Quantification of urban heat island phenomenon with remote sensing and machine learning techniques abstract: 背景城市熱島（UHI）是研究城市熱環境中不可或缺的主題，近年來越來越受到關注。然而，對於影響地表熱島（SUHI）空間變異的基本機制和驅動因素的研究，以及台灣這樣複雜地形區域中SUHI與樹冠層熱島（CLHI）之間的比較尚未有報告。因此，本研究旨在實現三個主要目標：（1）利用MODIS的地表溫度（LST）數據，探索選定的十一個台灣城市的SUHI強度（SUHII）及其相關決定因素的日夜、季節和空間特性；（2）利用機器學習（ML）模型，在2003年至2020年間，生成一個具有1公里分辨率的台灣月均空氣溫度（Ta）數據集（涵蓋四個衛星過境時間和每日均值）；以及（3）對SUHII和CLHI強度（CLHII）之間的時空模式進行比較分析。方法論這項工作經歷了四個主要的處理階段，包括城市區域的定義（第一階段）、SUHI特性和驅動因素的調查（第二階段）、使用ML算法估算Ta（第三階段）以及比較SUHI和CLHI（第四階段），以實現研究目標。更具體地，在第一階段中，使用隨機森林算法生成的ISA地圖首先利用來自LULC地圖和移動窗口技術的建成區域建立。接著，本研究使用ISA百分比、水類別和海拔數據來定義不同的區域。在第二階段中，使用MODIS LST數據和精煉的城市區域，計算了十一個台灣城市的UHI強度。接著，採用皮爾遜相關方法和逐步多元迴歸來研究與SUHII時空模式相關的因素。在第三階段，首先為台灣開發了月均Ta數據集（每日平均和四個過境時間），涵蓋了2003年至2020年的時期。這些數據集的空間分辨率為1公里，利用整合一系列變量的ML模型進行開發，包括現場測量、遙測資料和輔助信息。在第四階段，首先利用第三階段創建的Ta數據集計算了CLHI。然後，本研究比較了SUHII和CLHII的三個方面：（1）在日、季節或年度尺度上的強度差異；（2）利用SPAtial EFficiency（SPAEF）的空間分布差異；以及（3）利用皮爾遜相關分析研究SUHII和CLHII之間的關係。結果這項調查的主要發現包括： (1) 這項研究針對地表熱島強度（SUHII）及其相關決定因素的日夜、季節和空間特性進行了探討，發現SUHII在白天通常比晚上更高。此外，台灣北部城市的SUHII強度比南部城市高。SUHII在季節間有顯著波動，白天的變異性比晚上更大。白天的SUHII空間分佈受到歸一化差異潛熱指數（NDLI）、植被活動、建築密度（BI）和人為排放等因素的顯著影響。相反，夜間的SUHII與BI、夜間光信號和植被活動密切相關。這項研究的結果顯示，考慮的因素在白天SUHII變化中佔有更大比例，暗示夜間SUHII的潛在機制更為複雜。 (2) 就使用ML模型估算Ta而言，極端梯度提升（XGB）迴歸器表現優於其他兩個模型，呈現最高水準的準確度。該模型通過遞歸特徵消除（RFE）識別了八個關鍵變量。方差重要性分析的結果凸顯了夜間地表溫度（LST）做為最關鍵的預測因子，其次是白天LST和月份。由此產生的月均Ta數據集在5倍交叉驗證（5-CV）設置中表現良好（R2為0.986，RMSE為0.639°C，MAE為0.477°C）。值得注意的是，XGB模型在所有四個季節中始終表現出色，並在不同時間範圍（包括月份、年份和子集）下展現了顯著的準確性。此外，本研究的XGB模型在5-CV中取得了優秀的表現，分別為0.981（0.57）、0.980（0.60）、0.986（0.46）和0.986（0.45°C）在10:30、13:30、22:30和01:30時。 (3) 在比較兩種類型的城市熱島時，本研究觀察到全城市的年平均SUHII白天高於夜晚，但CLHII則呈現相反的模式。從季節性的角度來看，夜間SUHII-CLHII差異呈現出較弱的季節性變化。相反地，白天SUHII-CLHII差異觀察到顯著的季節變化，分別在溫暖季節具有較高值，在寒冷季節則具有較低值。至於空間分佈比較，夜間SUHI與CLHI相對相似（高SPAEF），但白天則有較大的差異（低SPAEF）。此外，相關性分析顯示各城市之間的SUHII與CLHII存在顯著且正向的相關性。結論本研究的發現為台灣的城市熱島提供了重要信息，例如SUHI的時空模式及其驅動因素，以及CLHI和SUHII之間強度和空間分佈差異的時空變化。因此，這項工作為發展SUHI預測和熱輻射緩解策略奠定了科學基礎。此外，本研究指出，具有高準確性的月均Ta數據集（四個過境時間和每日平均）可以成為環境研究和各種應用的重要因素。;Background The urban heat island (UHI), an essential topic in studying urban thermal environments, has drawn increasing attention in recent years. However, research on underlying mechanisms and drivers influencing the spatial variation of surface UHI (SUHI) and the comparisons between SUHI and canopy layer heat island (CLHI) for a complex topographic region like Taiwan have yet to be reported. Therefore, this study aims to achieve three primary objectives: (1) to explore the diurnal, seasonal, and spatial characteristics of SUHI intensity (SUHII) and its associated determinants across eleven chosen Taiwanese cities using land surface temperature (LST) data from MODIS; (2) to generate a monthly air temperature (Ta) dataset (at four overpass times and daily mean) with a resolution of 1 km employing machine learning (ML) models covering Taiwan from 2003 to 2020; and (3) to perform a comparative analysis of the spatiotemporal patterns between the SUHII and CLHI intensity (CLHII). Methodology This work underwent four main stages of processing, including the definition of the urban zones (stage 1), investigation of characteristics and driving factors of SUHI (stage 2), Ta estimation using ML algorithms (stage 3), and comparison between SUHI and CLHI (stage 4), to fulfill the research objectives. More specifically, in stage 1, the ISA map generated using the Random Forest algorithm was first created using the built-up area from the LULC map and the moving window technique. Next, this study used ISA percentage, water class, and elevation data to define different zones. In stage 2, UHI intensity was computed for eleven Taiwanese cities using the MODIS LST data and the refined urban zones. Next, Pearson′s correlation method and stepwise multiple regression were utilized to investigate the factors associated with the SUHII spatiotemporal patterns. In stage 3, monthly Ta datasets (daily average and four overpass times) were first developed for Taiwan, covering the period from 2003 to 2020. These datasets were developed with a spatial resolution of 1 km, utilizing ML models that integrated a range of variables, including in situ measurements, remotely sensed data, and auxiliary information. In stage 4, first, the Ta datasets created in stage 3 were used to compute the CLHI. Then, this study compared SUHII and CLHII in three aspects: (1) the differences in intensities at diurnal, seasonal, or annual scales; (2) differences in spatial distribution using the SPAtial EFficiency (SPAEF) and (3) the relationships between SUHII and CLHII employing Pearson′s correlation analysis. Results The main findings of this investigation include: (1) As for the diurnal, seasonal, and spatial aspects of SUHII and its related determinants, this work′s findings reveal that SUHII tends to be greater during the daytime than at nighttime. Furthermore, northern cities of Taiwan exhibit higher SUHII intensities than southern cities. The SUHII experiences notable seasonal fluctuations, with more significant variability observed during the daytime than at nighttime. The daytime spatial distribution of SUHII is significantly influenced by factors such as the Normalized Difference Latent Heat Index (NDLI), vegetative activity, built-up intensity (BI), and human-induced emission. Conversely, nighttime SUHII is closely associated with BI, nighttime light signal, and vegetative activity. The findings from this research indicate that the factors considered explain a more significant portion of the daytime SUHII variation compared to nighttime, indicating the more complicated underlying mechanism for nighttime SUHII. (2) As for the Ta estimation using ML models, the Extreme Gradient Boosting (XGB) regressor outperformed the other two models, demonstrating the highest level of accuracy. The model identified eight key variables through Recursive Feature Elimination (RFE). Results from variance importance analysis highlighted nighttime LST as the most critical predictor, subsequently daytime LST and month. The resultant monthly Ta dataset performed satisfactorily in a 5-fold cross-validation (5-CV) setting (R2 of 0.986, RMSE of 0.639 °C, and MAE of 0.477 °C). Notably, the XGB model consistently excelled across all four seasons and demonstrated remarkable accuracy over various time frames, including months, years, and subsets. Furthermore, this study′s XGB model yielded outstanding 5-CV performances, with R2 (MAE) of 0.981 (0.57), 0.980 (0.60), 0.986 (0.46), and 0.986 (0.45°C) at 10:30, 13:30, 22:30, and 01:30, respectively. (3) As for the comparison of two kinds of UHIs, while this study observed a higher annual SUHII in the day than at night over all cities, an opposite pattern was found for CLHII. From a seasonal perspective, the nighttime SUHII-CLHII difference shows weak seasonal variation. In contrast, significant seasonal variations were observed for the daytime SUHII-CLHII difference, with high values for warm and low values for cold seasons, respectively. As for spatial distribution comparison, the SUHIs are relatively similar to CLHIs at night (high SPAEF) but more different in the daytime (low SPAEF). Besides, correlation analysis shows significant and positive correlations between SUHII and CLHII across cities. Conclusion This study′s findings provide crucial information on urban heat islands in Taiwan, such as the spatiotemporal patterns of SUHI and its driving forces and the spatiotemporal variations of the difference between CLHI and SUHII in terms of intensities and spatial distribution. Thus, this work formed a scientific basis for developing SUHI prediction and heat mitigation strategies. Also, this study suggests that the monthly Ta datasets (four overpass times and daily mean) with high accuracy can serve as a crucial factor in environmental research and diverse applications.