| 摘要: | 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. |
