隨著全球暖化與氣候變遷加劇，紅樹林作為關鍵的「藍碳」生態 系，其保育與復育工作對於減緩溫室效應及維護生物多樣性具有重要意 義。然而，傳統的紅樹林監測方法多依賴人工調查，成本高昂且難以進 行大範圍的動態追蹤。儘管現有遙測技術已能有效偵測紅樹林的退化 （Loss），但對於紅樹林復育（Gain）與再生狀況的精確評估仍面臨諸多 挑戰，特別是在區分自然生長與人為種植的細微變化上。 本研究提出了一套基於GoogleEarthEngine (GEE) 與深度學習的自 動化監測框架，旨在以低成本、高效率的方式偵測紅樹林的面積變化， 並特別著重於復育成效的評估。研究資料選用Sentinel-2光學衛星影像， 涵蓋佛羅里達（Florida）與巴基斯坦（Pakistan）兩大研究區域，以驗證 模型在不同氣候與地理環境下的泛化能力。在方法論上，本研究首先利 用GEE平台進行大數據前處理，將影像從大氣頂層反射率（TOA）校正 為大氣底層反射率（BOA），並合成無雲影像。接著，本研究提出了一 種創新的時空特徵融合網路架構。該模型整合了改進型Inception模組與 卷積長短期記憶網路（ConvLSTM），並引入自研發的差分特徵編碼模組 （DFEM）以強化對微小變遷區域的特徵提取。此外，透過增強型殘差細 化單元（ERRB）與混合損失函數（FocalLoss+DiceLoss）的應用，有 效克服了紅樹林復育偵測中嚴重的正負樣本不平衡問題。 實驗結果顯示，本研究提出的模型能有效解決紅樹林變遷偵測中常見的樣本不平衡問題，不僅在偵測紅樹林消失方面表現優異，在識別紅 樹林復育區域上也展現了高度的準確性。本研究成果預期能為全球紅樹 林保育政策提供科學依據，並證明深度學習技術在生態復育監測上的應 用潛力。;With the intensification of global warming and climate change， mangroves， as a key ”blue carbon” ecosystem， are crucial for mitigating the greenhouse effect and maintaining biodiversity through conservation and restoration efforts. However， traditional mangrove monitoring methods largely rely on manual surveys， which are costly and difficult to conduct large-scale dynamic tracking. Although existing remote sensing technologies can effectively detect mangrove loss， accurate assessment of mangrove gain and regeneration remains challeng ing， particularly in distinguishing subtle changes between natural growth and human planting. This study proposes an automated monitoring framework based on Google Earth Engine (GEE) and deep learning， aiming to detect mangrove area changes in a low-cost and efficient manner， with a particular focus on assessing restora tion effectiveness. Sentinel-2 optical satellite imagery covering two major study areas—Florida and Pakistan—was used to validate the model’s generalization ability under different climatic and geographical conditions. Methodologically，this study first utilizes the GEE platform for big data preprocessing， correcting the imagery from Top Atmospheric Reflectance (TOA) to Bottom Atmospheric Reflectance (BOA) and synthesizing cloud-free imagery. Next， this study proposes an innovative spatiotemporal feature fusion network architecture. This model integrates an improved Inception module with a Convolutional Long Short-Term Memory (ConvLSTM) network and introduces a selfdeveloped Differential Feature Encoding (DFEM) module to enhance feature extraction from regions of subtle changes. Furthermore， the application of En hanced Residual Refinement Units (ERRB) and a hybrid loss function (Focal Loss + Dice Loss) effectively overcomes the severe positive-negative sample imbalance problem in mangrove restoration detection. Experimental results show that the proposed model effectively solves the common sample imbalance problem in mangrove changedetection， demonstrating excellent performance not only in detecting mangrove disappearance but also in identifying mangrove restoration areas with high accuracy. The findings of this research are expected to provide a scientific basis for global mangrove conservation policies and demonstrate the application potential of deep learning technology in ecological restoration monitoring.
