利用類神經網路方法於高解析衛星影像及地形資料之崩塌地辨識_以九份二山為例

Landslide Features Interpreted by Neural Network Method Using a High-Resolution Satellite Image and Terrain Data-A Joer-Fen-Ell Mountain Case

張崑宗(K. T. Chang) ； 劉進金(J. K. Liu)

《航測及遙測學刊》 11卷2期 (2006/06) Pp. 161-174

https://doi.org/10.6574/JPRS.2006.11(2).4

摘要

崩塌地是地球表面為求得動態平衡所引發的自然現象。台灣多颱風及地震，基於永續經營目的，台灣水庫管理單位每五到十年就對集水區範圍內崩塌地做一全面調查。自七零年代起，航照人工判釋技術被認為是崩塌地調查上最可靠的方法。本文中，將人工判釋之經驗法則加以量化，結合高解析衛星影像、光達所獲得DEM資料、以及道路及河流等向量資料，求取辨識特徵，以統計因子分析方法了解各特徵問關聯性。接著，利用多層次認知類神經網路進行崩塌地辨識訓練與測試。最後將所得結果與人工判讀結果比較與評估。結果顯示代表裸露地表之色調因子於崩塌地辨識上為一種顯著特徵。崩塌地與非崩塌地辨識精度最高分別可達98%與88%，可知本文所提出方法可有效輔助崩塌地調查。

關鍵字

類神經網路 ； 高解析力 ； 衛星影像 ； 空載光達 ； 崩塌地

並列摘要

Landslides are natural phenomena for the dynamic balance of earth surface. Due to the frequent occurrences of Typhoons and earthquake activities in Taiwan, mass movements are common threatens to our lives. Moreover, it is a common practice for the agencies of water reservoirs in Taiwan to make a reconnaissance of the landslides of the watershed every 5 to 10 years for the purpose of conservation. It is found that the application of aerial photo-interpretation technique for this purpose has been recognized as an effective approach since 1970s. However, an efficient and automatic interpretation scheme has never been established. Therefore, two issues are to be resolved for creating a useful and timely landslide database, i.e. the consistency of the sub-datasets and the completeness of the coverage. As the manual interpretation and automatic recognition are compared, the former is a practical and operational method, but the result it derived is largely dependent on the professional background of interpretation operator.
In this paper, the interpretation knowledge is quantified into recognition criteria. Multi-source data, e.g. a Quickbird satellite image, DTM reduced from a LIDAR data, road and river vector data, are fused to construct the feature space for landslides analysis. Then, those features are used to recognize landslides by a multilayer perceptron (MLP) Neural Network Method. The extraction result is evaluated in comparison with the manual-interpretation result. The experiments indicate that the bare surface (color tone factor) is a significant interpretation key for landslides recognition. In this case study, the highest recognition accuracy for landslides and non-landslides are 98% and 88%, respectively. Therefore, the conducted method can assist landslides investigation efficiently and automatically.

並列關鍵字

Artificial Neural Networks ； high-resolution ； satellite image ； airborne LIDAR ； Landslides