| 摘要: | 山洪對越南山區的生命、基礎設施和財產構成嚴重且不斷升級的威脅。 二十年來，
越南社會經濟快速發展，成為地區發展最快的國家之一。 然而，這一進步伴隨著城
市化、土地覆蓋變化和天然森林覆蓋率的減少，加劇了山洪爆發的風險。
本研究重點在於兩個目標：（1）應用機器學習（ML）模型預測山洪爆發敏感性；
（2）評估近二十年來人類活動對越南山區山洪暴發敏感度的定量影響。
方法
為了實現上述兩個目標，研究檢視了452個歷史山洪點，並分析了15個獨立因素，
包括海拔、坡度、坡向、曲率、地形濕度指數（TWI）、河流功率指數（SPI）、
流量累積、河流密度、到河流的距離、NDVI、NDBI、土地利用/土地覆蓋 (LULC)
。根據這些數據，研究採用了各種機器學習演算法來預測山洪爆發的機率，包括邏
輯回歸（LR）、K最近鄰（KNN）、高斯樸素貝葉斯（GNB）、多層感知器（MLP）
、支援向量機 (SVM)、隨機森林 (RF) 和 XGBoost (XGB)。使用表現出最高性能的
演算法來建立 2001-2010 年和 2013-2022 年期間的山洪暴發敏感度圖。 為了評估人
為影響，我們對土地利用模式的變化進行了詳細分析，並採用了歸一化植被指數（
NDVI）和歸一化差異建成指數（NDBI）等指數。 這些因素在塑造兩個時期山洪發
生機率的差異方面發揮了重要作用。
iii
結果
結果表明，RF、XGB 和 KNN 模型表現出卓越的性能，準確率分別為 0.949、0.941
和 0.919，以及令人印象深刻的 AUC-ROC 值（分別為 0.990、0.989 和 0.974）。
研究還表明，過去二十年人類社會經濟發展活動使高易發區和極高易發區發生山洪
的機率分別增加了7.69%和4.01%，令人擔憂。
結論
這項開創性的努力引入了一套新穎且全面的關聯模型，為現有的洪水預測方法增添
了重要價值。 此外，這些發現提供了切實而有力的證據，決策者可以利用這些證據
來評估持續的社會經濟成長的影響。 此外，它們是製定永續發展計畫的重要基礎，
該計畫優先考慮減輕未來不斷升級的山洪風險。;
Background
Flash floods pose a significant and escalating threat to life, infrastructure, and property in
the mountainous regions of Vietnam. Over the past two decades, Vietnam has experienced
rapid socio-economic development, making it one of the fastest-growing countries in the
region. However, this progress has been accompanied by urbanization, land cover
conversion, and reductions in natural forest coverage, exacerbating the risk of flash floods.
This study focuses on two objectives: (1) Applying the Machine learning (ML) model to
predict the flash flood susceptibility and (2) Evaluating the quantitative influence of human
activities on flash flood susceptibility in recent two decades in mountainous of Vietnam.
Methodology
To solve the above two objectives, the study has examined 452 historical flash flood
points and analyzed 15 independent factors encompassing elevation, slope, aspect, curvature,
topographic wetness index (TWI), stream power index (SPI), flow accumulation, river
density, distance to the river, NDVI, NDBI, land use/ land cover (LULC), rainfall, soil type,
geology. From these data, the study has employed various machine learning algorithms to
predict the probability of flash flood, including Logistic Regression (LR), K-Nearest
Neighbors (KNN), Gaussian Naïve Bayes (GNB), Multi-layer Perceptron (MLP), Support
Vector Machines (SVM), Random Forests (RF), and XGBoost (XGB). The algorithm that
exhibits the highest performance was used to build the flash flood susceptibility maps for the
period of 2001-2010 and 2013-2022. To assess the anthropogenic impact, we conducted a
detailed analysis of changes in land use patterns and employed indices such as the
v
Normalized Difference Vegetation Index (NDVI) and Normalized Difference Built-up Index
(NDBI). These factors played a significant role in shaping the differences in flash flood
probability between the two time periods.
Results
The results demonstrated that RF, XGB and KNN models showcasing superior
performance, boasting accuracy rates of 0.949, 0.941, and 0.919, along with impressive
AUC-ROC values of 0.990, 0.989, and 0.974, respectively.
The study also showed that human socio-economic development activities in the last two
decades have increased alarmingly in the probability of flash floods by 7.69% and 4.01% in
areas classified as high and very high susceptibility, respectively.
Conclusion
This pioneering endeavor introduces a novel and comprehensive suite of associative
models, adding significant value to existing flood prediction methodologies. Besides, these
findings provide tangible and robust evidence that policymakers can utilize to evaluate the
implications of ongoing socio-economic growth. Furthermore, they serve as a critical
foundation for formulating sustainable development plans that prioritize mitigating the
future escalating risk of flash floods. |
