深度學習影像分類應用於福衛二號衛星影像之崩塌地自動判釋

Abstract

臺灣易遭受各種自然災害侵襲，如地震或颱風。山區容易發生崩塌，造成生命財產安全的損害。使用遙測技術監測崩塌地的發生是政府機關的年度任務。然而，這個任務暨需求大量人力，且花費許多時間。為了解決這個問題，本研究提出應用深度學習進行全自動化衛星影像崩塌地分類的方法，以獲得更準確並可靠的分類結果。所採用之分類模型是基於U-Net 卷積類神經網路，以CNTK 深度學習工具進行開發。此模型以一對衛星影像與地真資料做為模型輸入，輸出預測的分類結果。使用多對福衛二號影像與地真資料進行模型訓練。地真資料共分五類：植被、河床、崩塌地、水體與其他地物類別。為了更準確分辨崩塌地與其他易混淆類別如河床與農地，衛星影像中加入坡度圖層做為分類資訊。本研究所產出的分類模型相當可靠，能夠清楚從衛星影像上判釋崩塌地。模型本身可以重複使用，而且過程完全自動化，並可改善崩塌地監測與繪製崩塌地目錄的工作流程，對環境災害監測與災害潛勢繪製將有所助益。

Taiwan is subject to severe natural hazards like earthquakes and typhoons, which often cause landslides in mountainous area, destroying crops and properties or even lives. Monitoring the occurrence of landslidesusing remote sensing imagery is an annual task for government institutions. However, the task had beenextremely labor-intensive and time consuming. In order to solve the problem, this study proposes a deeplearning technique for automatic landslide classification from satellite imagery in order to get more accurateand robust classification results.The classification model is based on the U-Net convolutional neural network. The model takes pairs ofsatellite imagery and ground truth label as the input and produces predicted classified labels as the output.The model is trained on pairs of FORMOSAT-2 imagery and ground truth labels. The ground truth isclassified into 5 classes: vegetation, riverbed, landslides, water and miscellaneous. To best separate landslidesfrom other unclear land cover like riverbed and farmlands, slope degree is added to satellite imagery todistinguish and recognize information for classification.The study’s results produce a robust classification model that is able to distinguish landslides from thesatellite imagery with an automatic workflow. We expect that the model will be useful for landslides monitoringand inventory mapping, which are elementary tasks for hazard mitigation and susceptibility mapping.