DSpace community: 研究中心

A multi-source remote sensing and Earth system model projection framework for quantifying climate change impacts on rice productivity across the Mekong River Delta;A multi-source remote sensing and Earth system model projection framework for quantifying climate change impacts on rice productivity across the Mekong River Delta

title: A multi-source remote sensing and Earth system model projection framework for quantifying climate change impacts on rice productivity across the Mekong River Delta;A multi-source remote sensing and Earth system model projection framework for quantifying climate change impacts on rice productivity across the Mekong River Delta abstract: 湄公河三角洲（Mekong River Delta， MRD）生產越南半數以上的稻米，並支撐全國主要的稻米出口，為東南亞重要的糧食生產區。然而，其地勢低平且高度依賴水文與氣候條件，使其對氣候變遷所引發的高溫、降雨異常、乾旱、鹽分入侵與洪水等複合性衝擊高度敏感。傳統氣候脆弱度評估多仰賴間歇性統計資料及空間分布有限的地面氣象觀測，不僅難以刻畫三角洲內部細緻的空間異質性，亦難以將歷史觀測所得之關係延伸至超出歷史變異範圍的未來氣候情境。因此，如何整合具空間連續性之觀測資訊與具情境推估能力之氣候模式資料，以量化不同氣候變遷情境下稻米生產力的時空變化，為評估湄公河三角洲未來糧食安全與氣候調適能力的重要科學課題。本研究針對上述「觀測—推估」之間的尺度與資訊落差，建立一套以遙測為核心、整合多源衛星觀測與氣候模式推估資料之應用分析架構，系統性評估不同排放情境下，湄公河三角洲各省份及不同稻作季節之氣候變遷衝擊，並進一步推估至2075年之稻米生產力變化。研究架構首先利用多源遙測資料建立歷史基線條件：以MODIS時間序列資料反演稻作物候期，以諧合化（harmonized）Landsat與Sentinel-2影像監測植被狀態與土地鹽分變化；利用Sentinel-1合成孔徑雷達（Synthetic Aperture Radar， SAR）建立歷年農地淹水資訊，並結合年度土地覆蓋資料估算稻作耕作面積。未來氣候條件則採用1995—2075年六組NEX-GDDP-CMIP6全球氣候模式系集資料，建立不同共享社會經濟路徑（Shared Socioeconomic Pathways， SSPs）排放情境下之氣候變遷推估。在稻作氣候脅迫量化方面，本研究依據遙測反演之稻作物候期，建立五項與稻米生產力密切相關之氣候脅迫指標，包括生長早期強降雨、結穗期強降雨、生殖期熱脅迫、乾旱—鹽分入侵脅迫，以及上游驅動之洪水訊號，以刻畫不同生育階段所承受之氣候風險。針對湄公河流域上游降雨與三角洲下游淹水之空間與時間傳遞關係，進一步利用對數線性轉換函數建立上游強降水與下游淹水範圍之關聯模型。為改善氣候模式與降雨產品空間解析度不足及系統性偏差問題，本研究另發展兩階段FNN–UNet降尺度模型，將每日降水資料降尺度至1 km空間解析度，以提升降雨分布估計的空間細緻度及其對局地氣候脅迫的表徵能力。在產量推估方面，本研究首先利用歷史統計資料分離稻米生產之長期趨勢與氣候變異效應，進而結合線性迴歸與前饋神經網路（Feedforward Neural Network， FNN），以所建立之氣候脅迫指標作為預測因子，分別推估稻米單位面積產量與總產量，藉以建立從「氣候變化—氣候脅迫—稻作面積—稻作產量」之量化鏈結。歷史分析結果顯示，1995—2024年間湄公河三角洲稻米產量的成長主要來自單位面積產量提升，其中約91%的產量增長可歸因於單位產量的增加，顯示過去三十年間生產效率提升為區域稻米增產的主要驅動因素。然而，未來氣候推估顯示，熱脅迫將逐漸成為影響稻作生產力的主要氣候暴露因子。其中，生殖期熱脅迫呈現最為明顯的增長趨勢，其關鍵暴露日數約為每年27天；至本世紀末，於中度排放情境下將增加至約68天，高度排放情境下更可能達到約107天，顯示稻作生殖階段所面臨的高溫風險將顯著加劇。另一方面，研究亦發現內陸省份之乾旱條件大致維持於現有水平，而沿海省份的降雨不足則呈持續加深趨勢，反映三角洲不同區域所承受的氣候風險具有顯著空間異質性。就整體產量而言，湄公河三角洲稻米總產量預估將於2035年前後達到約2，750萬公噸的高峰，隨後逐步下降，至2075年約為2，650萬公噸。值得注意的是，中度（SSP2-4.5）與高度排放情境（SSP2-8.5）下的總產量差異並不顯著，顯示單純以三角洲總量衡量氣候變遷衝擊，可能低估其在空間與季節尺度上的結構性變化。進一步分析各省份及不同稻作季節後發現，總產量相對穩定的表象下存在明顯的空間與季節重分配現象：夏秋作於多數省份呈現產量下降，其中以西南沿海地區最為顯著，包括金甌、薄寮與朔莊；相較之下，內陸地區冬春作之產量則大致維持穩定。此結果顯示，氣候變遷對湄公河三角洲稻作生產的影響並非均勻發生，而是受到地理位置、稻作季節與局地水文環境共同調控。在各項氣候脅迫因子中，上游洪水訊號與強降水對稻米單位面積產量呈現最明確的負向影響，顯示稻作生產不僅受到局地氣候條件控制，亦受到流域尺度水文傳遞所形成之跨區域洪水風險影響。進一步的空間分析顯示，各類氣候風險具有明顯的空間聚集特性：內陸淡水核心區整體具有較高的氣候韌性；相對而言，沿海地區同時暴露於高溫與鹽分入侵等複合性脅迫，並伴隨約12%的稻作耕作面積縮減，使沿海地區成為湄公河三角洲稻米生產系統中氣候暴露程度最高且最具脆弱性的區域。本研究之核心貢獻在於建立一套以多源遙測觀測為基礎、結合氣候模式推估、深度學習降尺度與產量機制分析之整合性評估架構。相較於傳統依賴稀疏地面觀測與歷史統計資料的氣候脆弱度評估方法，本研究透過目前公開可取得之衛星遙測資料，建立具空間連續性與高解析度的稻作物候、植被、鹽分、淹水及耕作面積資訊，並將其與未來氣候情境及稻米產量模型相互串聯，從而突破歷史觀測資料在空間覆蓋與未來情境推估上的限制。研究結果進一步揭示，氣候變遷對湄公河三角洲稻米生產的主要影響並非單純表現為總產量下降，而是透過高溫、降雨、洪水、乾旱、鹽分入侵及耕作面積變化等多重因素，造成不同省份與稻作季節間的生產力重新分配。因此，本研究所提出之架構可提供具空間明確性、時間連續性及情境推估能力之科學資訊，作為湄公河三角洲糧食安全評估、農業生產管理與氣候變遷調適規劃之重要決策支援工具，並為大尺度氣候變遷衝擊評估提供一套具可操作性與可擴充性的遙測整合分析方法。;The Mekong River Delta (MRD) produces more than half of Vietnam’s rice and accounts for the majority of the country’s rice exports. Its low-lying terrain， however， makes the delta particularly vulnerable to climate change and associated hydroclimatic extremes. Conventional vulnerability assessments largely rely on lagged statistical records and sparsely distributed weather stations， which are insufficient to resolve the pronounced spatial heterogeneity of climate exposure across the delta and provide limited capacity to project impacts beyond the range of historical observations. To address this observation–projection gap， this study develops a transferable， remote-sensing-driven framework that integrates multi-source satellite observations with climate projections to quantify climate-change impacts on rice productivity across provinces and cropping seasons in the MRD under moderate- and high-emission scenarios through 2075. The framework establishes a consistent linkage between observed agricultural conditions and projected future climate stressors. Historical baseline conditions are characterized using multi-source satellite remote sensing， including MODIS-derived rice phenology， harmonized Landsat and Sentinel-2 observations of vegetation and salinity stress， Sentinel-1 synthetic aperture radar (SAR) observations of surface flooding， and annual land-cover maps to quantify cultivated area. Future climate conditions are derived from a six-member NEX-GDDP-CMIP6 ensemble spanning 1995–2075. Satellite-derived phenological windows are used to constrain five crop-specific climate stress indicators: heavy rainfall during the early growth stage， heavy rainfall around flowering， reproductive-stage heat stress， a drought–salinity indicator， and an upstream-driven flood signal. To represent the propagation of hydroclimatic forcing through the river system， a log-linear transfer function is developed to relate upstream heavy precipitation to downstream flood extent. In parallel， a two-stage feedforward neural network–U-Net (FNN–UNet) model downscales daily precipitation to 1-km resolution， thereby improving the spatial representation of precipitation and reducing systematic biases associated with coarse-resolution products. For production attribution and projection， long-term yield trends are first separated from interannual climate variability using historical production records. The resulting detrended yield anomalies are then modeled using a hybrid linear-regression and feedforward neural network (FNN) approach driven by the climate-stress indicators. This formulation provides a quantitative linkage between projected climate stressors and both yield response and total rice production， while allowing the relative influence of individual stressors to be examined. Historical decomposition indicates that approximately 91% of the increase in MRD rice production from 1995 to 2024 was attributable to rising yield rather than expansion of cultivated area， highlighting yield improvement as the dominant historical driver of production growth. Future projections under the Shared Socioeconomic Pathways (SSPs) indicate that reproductive-stage heat stress will become the dominant exposure signal. The number of critical heat-stress days is projected to increase from approximately 27 days per year historically to 68 days under SSP2-4.5 (middle of the road) and 107 days under SSP5-8.5 (fossil-fueled development) by the late twenty-first century. Meanwhile， drought conditions in the inland delta remain broadly near their historical range， whereas rainfall deficits intensify along the coastal belt， indicating increasingly divergent hydroclimatic trajectories across the region. At the delta-wide scale， however， the projected change in total production remains relatively modest. Production is projected to peak at approximately 27.5 million tons around 2035 before gradually declining to about 26.5 million tons by 2075， with limited separation between the moderate- and high-emission scenarios. This aggregate stability masks substantial spatial and seasonal redistribution. At the provincial and seasonal scales， Summer–Autumn rice production declines across most provinces， with the strongest reductions projected along the southwestern coastal belt， particularly in Ca Mau， Bac Lieu， and Soc Trang. In contrast， the inland Winter–Spring crop remains broadly stable or improves. The resulting pattern therefore represents a redistribution of production across seasons and locations rather than a uniform decline across the delta. Among the climate stressors， the upstream flood signal and heavy precipitation exhibit the clearest negative associations with detrended yield， providing the strongest evidence for stressor-specific attribution within the modeling framework. The spatial distribution of risk is likewise highly concentrated. The freshwater-dominated inland core retains comparatively high resilience， whereas the coastal belt experiences compound heat and salinity stress together with an approximately 12% reduction in cultivated area. Consequently， coastal rice production emerges as the most climate-exposed component of the MRD production system. Although the individual datasets and analytical components employed in this study are established， the principal contribution lies in their integration into a coherent， end-to-end assessment architecture. Built entirely upon openly accessible Earth observation data and coupled with climate-model projections， the framework links remotely sensed crop phenology， vegetation and salinity conditions， flood dynamics， cultivated area， climate-stress indicators， and production responses within a common spatially explicit framework. This integration addresses a fundamental challenge in climate-impact assessment: reconciling the spatial detail of observation with the forward-looking capacity of climate projection. The resulting province- and season-resolved evidence provides a scalable basis for identifying emerging climate hotspots， diagnosing the dominant drivers of production vulnerability， and supporting targeted food-security and climate-adaptation planning across the Mekong River Delta.

氣候極端與颱風影響下西北太平洋海洋熱浪與上層海洋增暖之時空變異;Spatiotemporal Variability of Marine Heatwaves and Upper-Ocean Warming Associated with Climate Extremes and Typhoon Influences in the Northwestern Pacific

title: 氣候極端與颱風影響下西北太平洋海洋熱浪與上層海洋增暖之時空變異;Spatiotemporal Variability of Marine Heatwaves and Upper-Ocean Warming Associated with Climate Extremes and Typhoon Influences in the Northwestern Pacific abstract: 西北太平洋為全球熱帶氣旋（tropical cyclone， TC；以下稱颱風）活動最頻繁之海域，近年海洋熱浪（marine heatwave， MHW）發生頻率與颱風破壞潛力均呈上升趨勢。本論文整合東海（ECS）、南海（SCS）及西太平洋（WP）三大海域之四項觀測研究，利用1993至2025年之衛星、再分析及最佳路徑資料，探討海表MHW究竟能為颱風增強提供實質能量，抑或僅為與次表層解耦之淺層熱異常。第一，海表MHW頻率與上層海洋熱含量（UOHC）及上層100公尺平均水溫（T100）於非颱風季呈中度相關（r = 0.4–0.6），惟於颱風季趨於解相關（熱帶WP與SCS海域 r→0）；La Niña與負太平洋年代際振盪（PDO）相位之複合氣候態，可將多層熱力極端事件放大約40%，超出單一相位效應之總和。第二，於91個東海颱風個案中，海表MHW之存在無法區分增強與非增強個案；即使內核UOHC偏高，增強機率亦僅約40%，主要受限於東海較淺之混合層（<25公尺）。第三，於1，027個颱風眼位置樣本中，MHW雖與較高颱風強度呈統計關聯，惟仍有151個個案（約15%）於活躍MHW環境下減弱，顯示TC–MHW耦合為非確定性過程。第四，於2，402個颱風通過事件中，颱風引致之海表降溫幅度隨強度等級遞增，南海較西太平洋高47%；猛烈颱風（≥54 m/s）通過後三日內，MHW嚴重度等級最多下降0.75級，其路徑右側之降溫不對稱與近慣性共振機制相符。綜上所述，海表MHW深度過淺，不足以供應颱風增強所需能量，且結構脆弱，難以於颱風通過後存續；次表層熱力結構較海表MHW狀態更適合作為颱風強度業務預報之海洋輸入因子。;The Northwestern Pacific， the world′s most active tropical cyclone (TC) basin， faces rising marine heatwave (MHW) frequency and destructive potential. This dissertation tests whether surface MHWs energize TC intensification or are shallow anomalies decoupled from the subsurface， through four studies across the East China Sea (ECS)， South China Sea (SCS)， and Western Pacific (WP) using 1993–2025 satellite， reanalysis， and best-track records. First， surface MHW frequency co-varies with upper-ocean heat content (UOHC) and mean upper 100 m temperature (T100) outside the TC season (r = 0.4–0.6) but decorrelates during it r→0， tropical WP and SCS)， while compound La Niña/negative Pacific Decadal Oscillation phases amplify multi-layer extremes ~40% beyond single-phase sums. Second， among 91 ECS TCs， surface MHW presence fails to discriminate intensifying from non-intensifying cases， and high inner-core UOHC yields only ~40% intensification probability， limited by the <25 m ECS mixed layer. Third， across 1，027 TC-eye positions， MHWs are statistically associated with greater intensity， yet 151 (~15%) TCs weaken under active MHWs， demonstrating non-deterministic coupling. Fourth， across 2，402 passages， TC-induced SST cooling scales with intensity class， is 47% greater in the SCS than in the WP， and suppresses MHW severity by up to 0.75 categories within three days of violent typhoon (≥54 m/s) passage， showing right-of-track asymmetry consistent with near-inertial resonance. Surface MHWs are too shallow to fuel intensification and too fragile to survive TC passage; subsurface thermal structure is a more appropriate predictor than surface MHW state for operational intensity forecasting.

基於深度學習之超高解析度衛星影像特徵萃取於大型底圖製作的應用;Deep Learning–Driven Feature Extraction from Very-High-Resolution Satellite Imagery for Large-Scale Basemap Production

title: 基於深度學習之超高解析度衛星影像特徵萃取於大型底圖製作的應用;Deep Learning–Driven Feature Extraction from Very-High-Resolution Satellite Imagery for Large-Scale Basemap Production abstract: 隨著印尼國家發展計畫的推行，對於大型底圖的需求顯著增加，而這些計畫需要精確且即時的空間資料。然而，現行的全國生產能力約為每年 13，000 平方公里，仍受限於依賴人力的立體測圖（stereo-plotting）及單像測圖（mono-plotting）流程，這些流程需使用航空攝影測量、光達（LiDAR）與高解析度衛星影像。為在維持製圖品質的同時加速大型底圖生產，本論文探討利用深度學習自動萃取超高解析度（VHR）衛星影像特徵的方法，並聚焦於兩個關鍵底圖圖層：土地覆蓋（land cover）與建築物輪廓（building footprints）。本研究首先對超高解析度衛星影像採用統一的特徵增強策略，以加強光譜、空間與紋理資訊，為後續學習任務提供更具辨識力的輸入。該策略包含全色態銳化（pan-sharpening）、標準化差異植生指數（NDVI）計算、灰度共生矩陣（GLCM）紋理分析，以及主成分分析（PCA）。在此共通特徵表示的基礎上，進一步建立深度學習模型，用於自動土地覆蓋分類與建築物輪廓擷取。在土地覆蓋製圖部分，本研究採用卷積神經網路（CNN）將 VHR 影像分類為符合國家大型地形圖製圖標準的土地類別。結果顯示，人工智慧輔助的土地覆蓋擷取可有效取代人工判讀，整體精度（Overall Accuracy）達 86%，Kappa 係數為 0.81，顯著提升製圖效率。在建築物輪廓生成方面，本研究使用經特徵增強的多波段影像訓練 Mask R-CNN 模型以分割建築物物件，並進行幾何正則化以簡化多邊形並使邊界對齊主要結構方向。以物件為基礎的評估結果顯示：精確率（Precision）97.6%、召回率（Recall）91.6%、F1 分數 94.5%。實務測試結果指出，該方法可將單張 5.29 平方公里地圖的生成時間縮短至一小時以內，相較傳統人工方法，效率提升超過 35 倍。整體而言，本論文證明利用深度學習自動萃取 VHR 衛星影像特徵，可提供精準、穩健且具可擴展性的解決方案，並適用於大型底圖製作。透過將自動土地覆蓋分類與建築物輪廓擷取整合於單一框架中，所提出的方法不僅大幅提升生產力，亦可滿足國家製圖、都市規劃及空間決策所需之品質要求。;The demand for large-scale basemaps in Indonesia has increased significantly following the implementation of national development programs that require accurate and up-to-date spatial data. However， the current national production capacity—approximately 13，000 km² per year—remains constrained by labor-intensive stereo-plotting and mono-plotting workflows using photogrammetric， LiDAR， and high-resolution satellite data. To accelerate large-scale basemap production while maintaining cartographic quality， this dissertation investigates the application of deep learning–driven feature extraction from very-high-resolution (VHR) satellite imagery， with a focus on two critical basemap layers: land cover and building footprints. A unified feature enhancement strategy is first applied to VHR satellite imagery to strengthen spectral， spatial， and textural information for downstream learning tasks. This strategy includes pan-sharpening， Normalized Difference Vegetation Index (NDVI) computation， Grey-Level Co-occurrence Matrix (GLCM)–based texture analysis， and principal component analysis (PCA). Building upon this common feature representation， deep learning models are then developed for automated land cover classification and building footprint extraction. For land cover mapping， convolutional neural networks are employed to classify VHR imagery into classes relevant to national large-scale topographic mapping standards. The results demonstrate that AI-assisted land cover extraction can effectively replace manual interpretation， achieving an Overall Accuracy of 86\% and a Kappa coefficient of 0.81， thereby significantly improving production efficiency. For building footprint generation， a Mask R-CNN model is trained on the enhanced multi-band imagery to segment building objects， followed by geometric regularization to simplify polygons and align edges with dominant structural orientations. Object-based evaluation against ground-truth data yields high accuracy， with 97.6% precision， 91.6% recall， and a 94.5% F1-score. Operational testing shows that the proposed approach reduces production time to less than one hour per 5.29 km² map sheet， representing more than a 35-fold efficiency gain compared to conventional manual methods. Overall， this dissertation demonstrates that deep learning–driven feature extraction from VHR satellite imagery provides a robust， accurate， and scalable solution for large-scale basemap production. By integrating automated land cover classification and building footprint extraction within a unified framework， the proposed methodologies significantly enhance productivity while satisfying quality requirements for national mapping， urban planning， and geospatial decision-making.

運用機器學習與多來源衛星及地理空間資料進行印度尼西亞廖內南部季節性野火易發性建模;Seasonal Wildfire Susceptibility Modelling Using Machine Learning with Multi-Source Satellite and Geospatial Data in Southern Riau， Indonesia

title: 運用機器學習與多來源衛星及地理空間資料進行印度尼西亞廖內南部季節性野火易發性建模;Seasonal Wildfire Susceptibility Modelling Using Machine Learning with Multi-Source Satellite and Geospatial Data in Southern Riau， Indonesia abstract: 印尼廖內南部的森林與土地火災具有明顯的季節差異，並受到植被、水分、氣象、土地利用及泥炭地條件等多重因素的共同影響。本研究建立全年、旱季及雨季的野火易發性模型，並評估歷史環境變化指標是否能提供額外的預測資訊。研究整合2010至2024年的多來源衛星與地理空間資料，並將其統一至500公尺網格；燒毀區標籤則由MODIS MCD64A1 BurnDate產品建立。當期條件型與條件變化型兩種預測因子配置，分別使用四種機器學習演算法進行比較。模型採用1:5的火災與非火災樣本比例，以及分層70:30訓練–測試資料切分，並以PR-AUC作為主要評估指標。最終選定的模型為全年與旱季的條件變化型LightGBM模型，以及雨季的當期條件型LightGBM模型；其保留測試集的PR-AUC分別為0.921、0.905及0.966。歷史環境變化指標僅小幅提升全年與旱季模型的表現，且未改善雨季模型，顯示其預測效益具有季節差異。2024年的分析結果顯示，全年與旱季易發性地圖具有較高的空間一致性，而雨季的高易發性區域則較為局部。旱季的高與極高易發性區域面積最大，達8，581平方公里，占分析網格總面積的21.49%。SHAP分析顯示，全年模型的主要影響因子為植被狀況、土壤水分趨勢及降水；旱季模型主要受到降水、植被狀況及油棕土地利用背景的影響；雨季模型則主要受到風速、植被狀況及泥炭地背景的影響。使用2024年印尼國家警察之獨立火災紀錄進行驗證的結果顯示，全年、旱季及雨季分別有77.60%、76.79%及78.57%的匹配紀錄位於高與極高易發性區域。研究結果顯示，分季建模能揭示僅使用全年模型時可能被掩蓋的預測差異與空間差異。;Forest and land fires in Southern Riau， Indonesia， exhibit strong seasonal variation and are influenced by multiple interacting factors， including vegetation， moisture， meteorological conditions， land use， and peatland characteristics. This study developed wildfire susceptibility models for the all-season， dry-season， and rainy-season periods and assessed whether historical environmental-change indicators provided additional predictive information. Multi-source satellite and geospatial data from 2010 to 2024 were harmonized to a 500 m grid， while burned-area labels were derived from the MODIS MCD64A1 BurnDate product. Condition-only and condition-change predictor configurations were compared using four machine-learning algorithms. The modelling design employed a 1:5 fire-to-non-fire sampling ratio and a stratified 70:30 train–test split， with PR-AUC as the primary evaluation metric. The selected models were condition-change LightGBM for the all-season and dry-season periods and condition-only LightGBM for the rainy-season period， with held-out test PR-AUC values of 0.921， 0.905， and 0.966， respectively. Historical environmental-change indicators produced only modest improvements in the all-season and dry-season models and did not improve the rainy-season model， indicating that their predictive benefit was season-dependent. The 2024 results showed stronger spatial agreement between the all-season and dry-season susceptibility maps， whereas high-susceptibility areas during the rainy season were more localized. The dry-season High and Very High classes covered the largest area， reaching 8，581 km²， or 21.49% of the analytical grid extent. SHAP analysis identified vegetation condition， soil-moisture trend， and precipitation as the principal contributors to the all-season model; precipitation， vegetation condition， and oil-palm land-use context as the principal contributors to the dry-season model; and wind speed， vegetation condition， and peatland context as the principal contributors to the rainy-season model. Independent validation using 2024 POLRI fire records showed that 77.60%， 76.79%， and 78.57% of the matched records for the all-season， dry-season， and rainy-season periods， respectively， occurred within the High and Very High susceptibility classes. These findings demonstrate that separate seasonal models can reveal predictive and spatial differences that may be obscured when wildfire susceptibility is represented solely by an all-season model.