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.
