從人工智慧圍棋程式「AlphaGo」屢屢戰勝棋王後，類神經網路以及深度學習技術成為眾所矚目的焦點。深度學習在遙感領域的應用日益廣泛，以目標物識別、土地覆被分類以及地表變異偵測等研究為主，利用機器學習工具發展不同影像來源之影像數據訓練，以做為進行廣域快速地表特徵變異之利器。隨著全球氣候變遷、極端災害事件頻傳，臺灣位於亞熱帶季風區、活躍的造山帶且地質材料破碎，經年累月受到颱風及地震等之自然災害侵襲，劇烈的地表作用造成大規模崩塌災害產生。為此，學術界以及相關主管機關長期投入坡地崩塌事件資料蒐集、致災因子評估與地質模型建置等工作，累積颱風、豪雨事件前後大量地表變異資料，提供深度學習在自動化崩塌判釋最佳之範例。本計畫擬以二年期間進行，以旗山溪集水區為研究範圍，整合ALOS雷達與福衛光學影像，利用卷積神經網路學習工具進行自動化崩塌判釋。第一年將以莫拉克災前後光達數值地形，搭配福衛二號影像與航空照片，進行潛在大規模崩塌地形特徵判釋。利用卷積神經網路工具進行研究區域已發生之大規模崩塌區位目錄，配合影響因子(差態化差異植生指標(NDVI)、岩體強度、坡度、坡向、高程、河道距離、構造距離、順向坡)，進行福衛2號大規模崩塌之影像數據訓練。第二年將以日本太空總署ALOS/PALSAR衛星雷達影像，利用衛星雷達雙偏極HH和HV影像強化目標物在不同偏極電磁波之辨識能力，配合前期福衛2號判釋成果進行地表變異分析。本研究預期成果包括:1)旗山溪流域福衛2號光學影像崩塌事件判釋成果與崩塌區位目錄;2)旗山溪流域ALOS雷達影像崩塌事件判釋成果與崩塌區位目錄;3)建立旗山溪流域歷次崩塌事件之卷積神經網路地表變異識別模式及圖像訓練資料庫。
After the AI-based Go program “AlphaGo” succeeded in beating a number of professional Go players, people started to pay great to the development of Convolutional Neural Network (CNN) and deep learning. In the field of remote sensing, deep learning has been widely applied with a focus on target recognition, land cover classification and surface change detection; and, to facilitate change detection on a large scale, machine learning tooling has been used with multiple imagery sources. Following climate change and the increasing number of extreme weather events around the world, Taiwan, which is situated in a subtropical monsoon climate region and active orogenic belt with fragile geological materials, frequently suffers from natural disasters such as typhoons and earthquakes. This has resulted in intense surface processes that lead to large-scale landslide hazards. Scholars and relevant competent authorities of the country therefore collect slope failure data, evaluate hazard-inducing factors and establish geological models as a long-term perspective. The enormous pre- and post-event displacement data collected thereby can then be offered to the deep learning system as training data to facilitate the development of automated landslide detection model. Targeting Chishan River watershed, this two-year project aims to develop an automated landslide detection model by integrating ALOS radar and Formosat-2 imagery, and feeding them to the convolutional neural network (CNN) as training data. In the first year, LiDAR derived DTMs (digital terrain models), Formosat-2 imagery and aerial photographs will be used to identify geomorphological features of deep-seated landslides; and large-scale landslides occurred in the study area will be inventoried using the CNN, which will be simultaneously fed with the inventory and landslide causative factors (ex. Normalized Difference Vegetation Index, NDVI; rock density; slope gradient; slope aspect; elevation; distance to river; distance to fault; and dip slope) as training data. In the second year, the interpretation results of prior Formosat-2 imagery and JAXA’s ALOS/PALSAR imagery will be jointly used to analyze the surface change of study area as dual-polarized (HH and HV) radar images can improve target recognition regardless the polarization of electromagnetic wave. The expected results of this project include: (1) identifying and inventorying landslides occurred in Chishan River watershed based on optical imagery of Formosat-2; (2) identifying and inventorying landslides occurred in Chishan River watershed based on ALOS imagery; (3) creating surface change recognition mode and imagery training datasets of CNN on the basis of large-scale landslides in Chishan River watershed.
