使用深度學習方法建立崩塌判釋模型

Abstract

崩塌災害對於居民造成財產和生命的威脅，同時也對下游水庫、河道輸砂和農業等方面造成影響。因此，若能快速且準確地掌握崩塌位置及其影響範圍，對於執行正確的減災策略和後續的整治工作至關重要。本研究使用深度學習方法對谷關水庫集水區進行崩塌判釋，並比較了不同輸入資料的模型表現差異，同時比較了深度學習方法與支持向量機的精度差異。研究結果顯示，深度學習方法進行崩塌判釋係具可行性，本研究使用衛星影像光譜值作為模型輸入資料，且加入不同崩塌影響因子，深度學習模型在測試集之F1-score介於0.75至0.81，精確率介於0.73至0.80，召回率介於0.77至0.85，整體表現良好。其中，當輸入資料加入平面曲率因子階段為最適選，F1-score為0.80，精確率為0.75，召回率為0.85，可成功判釋出85%衛星影像中崩塌類別。本研究亦比較深度學習方法和支持向量機模型表現。研究結果顯示，深度學習模型表現皆優於支持向量機模型。其中，於深度學習模型中加入平面曲率因子階段表現為最佳，F1-score為0.75，精確率為0.76，召回率為0.74，並於研究中發現深度學習模型判釋出的像素雜異點較少，不需額外處理雜異點問題，係使用深度學習模型優勢之一。Landslides pose threats to residents' lives and properties, as well as affecting downstream reservoirs, river sediment transport, and agriculture. Therefore, the ability to quickly and accurately identify the locations and extent of landslides is crucial for implementing effective disaster reduction strategies and subsequent recovery efforts. In this study, a deep learning approach was employed to interpret landslides in the Guanguan Reservoir watershed. The performance differences of models with various input data were compared,along with a comparison between the accuracy of deep learning and support vector machine (SVM) methods.The research findings demonstrate the feasibility of using deep learning for collapse interpretation. Spectral values from satellite images were used as input data for the model, along with various collapse influencing factors. The deep learning model achieved F1-scores ranging from 0.75 to 0.81, precision ranging from 0.73 to 0.80, and recall ranging from 0.77 to 0.85 on the test set, indicating overall good performance. Notably, the inclusion of the planar curvature factor as input data resulted in optimal performance, with an F1-score of 0.80, precision of 0.75, and recall of 0.85, successfully detecting 85% of landslide instances in satellite images.Furthermore, a comparison was made between the performance of the deep learning and SVM models. The results showed that the deep learning models outperformed the SVM models. Particularly, the model that incorporated the planar curvature factor demonstrated the best performance, with an F1-score of 0.75, precision of 0.76, and recall of 0.74. The study also revealed that the deep learning model produced fewer pixel anomalies, eliminating the need for additional anomaly handling. This underscores one of the advantages of using deep learning models.