以機器學習方法建立考量歷史崩塌及空間關聯性的崩塌潛勢預測模型

Abstract

台灣地處板塊交界帶，受造山運動及季風氣候影響，形成高強度侵蝕環境，崩塌發生頻繁，因此事前預警及防減災對坡地安全十分重要。隨地理資訊系統（GIS）和大數據發展，使探討地文、水文及人文等因子對坡地穩定的影響成為可能，其中透過機器學習建置崩塌潛勢評估模型是目前受矚目的方式。 傳統機器學習研究多以易取得的地形、地質、水文等資料為模型訓練特徵，但在崩塌頻發的台灣，已發生崩塌地區的穩定性及植被覆蓋相對較差，易再次崩塌。因此，本研究以歷史崩塌事件作為模型訓練特徵之一，並探討其重要性。另外，由於崩塌事件範圍不定、規模不一，網格切分易使相鄰網格相關性消失，影響模型評估準確性。因此，本研究引入自然語言轉譯領域的自注意力機制，自動化全局考量序列中各單元相關性，以解決網格間獨立性的問題。期許能夠建立廣範圍、高精度的崩塌潛勢評估模型。 結果顯示，加入歷史崩塌事件特徵能大幅提升模型評估能力，不論是否考慮空間相關性，都有明顯提升（F1_score分別於mlp模型及att模型提升0.38及0.51）。而自注意力方法使模型在precision及recall中取得較好平衡並降低資料偏誤造成之缺陷，更適用於實際防減災操作。在地質差異大的區域，本研究建立的評估模型保持良好的泛用性（F1_score維持在0.7左右）。評估不同年份事件，模型在大多數情況下保持高水準表現（F1_score維持在0.6以上），具有時間泛用性。與過往的研究相比，本研究取得更穩定及良好之崩塌評估表現。

Taiwan is located at the convergence zone of tectonic plates and influenced by orogeny and monsoon climate, which has evolved into an environment with intense erosion and frequent landslides. Therefore, proactive warning systems and disaster mitigation measures are of paramount importance for the slope safety. With the advancement of geographic information systems (GIS) and big data, the exploration of the impact of physical, hydrological, and anthropogenic factors on slope stability has become feasible. Among them, the construction of landslide susceptibility prediction models through machine learning stands out as a currently prominent approach. Conventional machine learning studies often utilize easily accessible topographical, geological, and hydrological data as features for the model training. However, in Taiwan, areas with historical landslide exhibit lower stability and vegetation cover, making them susceptible to recurring landslides. Hence, this study incorporated historical landslides as one of the model training features and assessed its importance. Furthermore, due to the variable extent and scale of landslides, grid partitioning can lead to the disappearance of correlation between neighboring grids, thereby affecting the accuracy of model predictions. Therefore, this research introduced a self-attention mechanism from the natural language processing domain to automatically and globally consider the correlation of each unit in a sequence to address the issue of grid independence. The goal is to establish a wide-ranging and high-precision landslide susceptibility prediction model. The results show that a substantial improvement in the model prediction after incorporating the historical landslides, having significant enhancements in both precision and recall, irrespective of spatial correlation considerations (F1_scores increased by 0.38 and 0.51 in the mlp and att models, respectively). The self-attention method enables the model to achieve a better balance between precision and recall, reducing defects induced by data biases, and is more suitable for practical disaster mitigation operations. In areas with significant geological variations, the prediction model established in this study maintains a good generalizability (F1_score consistently around 0.7). Predicting in different years, the model maintains high-level performance in most cases (F1_score consistently above 0.6), demonstrating the temporal generalizability. In comparison to previous studies, this research achieves a more stable and robust landslide prediction performance.