| 人工神经网络和决策树模型在滑坡易发性分析中的性能对比 |
| |
| 引用本文： | 田乃满,兰恒星,伍宇明,李郎平.人工神经网络和决策树模型在滑坡易发性分析中的性能对比[J].地球信息科学,2020,22(12):2304-2316. |
| |
| 作者姓名： | 田乃满 兰恒星 伍宇明 李郎平 |
| |
| 作者单位： | 1.中国科学院地理科学与资源研究所 资源与环境信息系统国家重点实验室,北京 1001012.长安大学地质工程与测绘学院,西安 7100543.中国科学院大学,北京 100049 |
| |
| 基金项目： | 中国科学院战略性先导科技专项(A类)(XDA23090301);国家自然科学基金项目(41701458);国家自然科学基金项目(41525010);国家自然科学基金项目(41790443);国家自然科学基金项目(41807291) |
| |
| 摘 要： | 机器学习模型广泛应用于区域性滑坡易发性分析。模型的选择关系到评价结果的可信度、准确率和稳定性。现有滑坡易发性分析模型对比研究侧重模型的预测精度。模型的稳定性和数据量敏感性对机器学习模型的性能评估同样非常重要。本文以福建省南平市蔡源流域为研究区,以四川省绵阳市北川县为验证区,从预测精度、稳定性和数据量敏感性3个方面深入对比BP（Back Propagation）人工神经网络模型和CART（Classification and Regression Tree）决策树模型在滑坡易发性分析中的效果,主要结论如下:① 在逐渐增加一定数量训练样本的过程中,BP人工神经网络模型预测精度的增长率更高。在蔡源流域内,当训练样本数量增加10 000时,BP人工神经网络模型的预测精度上升5.22%,CART决策树模型的预测精度上升2.11%。② BP人工神经网络的预测精度高于CART决策树模型,且较为稳定。在100组数据集上,BP人工神经网络模型验证集预测精度的均值和验证集滑坡样本预测精度的均值分别为81.60%和84.86%,高于CART决策树模型的72.97%和76.59%。与此同时,BP人工神经网络模型对应预测精度的标准差分别是0.32%和0.37%,小于CART决策树模型的0.35%和0.67%。③ BP人工神经网络模型分析的滑坡易发区相比CART决策树模型,更接近实际滑坡的空间分布。最后,北川县的验证实验也出现了相同的现象。
|
| 关 键 词： | 滑坡 易发性分析 人工神经网络模型 决策树模型 模型对比 预测精度 稳定性 数据量敏感性 |
| 收稿时间： | 2019-12-11 |
Performance Comparison of BP Artificial Neural Network and CART Decision Tree Model in Landslide Susceptibility Prediction |
| |
| TIAN Naiman,LAN Hengxing,WU Yuming,LI Langping.Performance Comparison of BP Artificial Neural Network and CART Decision Tree Model in Landslide Susceptibility Prediction[J].Geo-information Science,2020,22(12):2304-2316. |
| |
| Authors: | TIAN Naiman LAN Hengxing WU Yuming LI Langping |
| |
| Institution: | 1. State Key Laboratory of Resources and Environmental Information Systems, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China2. School of Geological Engineering and Geomatics, Chang'an University, Xi'an 710054, China3. University of Chinese Academy of Sciences, Beijing 100049, China |
| |
| Abstract: | Machine learning has been widely applied to analyze regional landslide susceptibility, such as the artificial neural network and decision tree model. Model selection depends on both the reliability and accuracy of model results, therefore comprehensively evaluating the performance of a model is necessary. Previous studies of landslide susceptibility focused more on the prediction accuracy of a model. However, model stability and model sensitivity to data volume also reflect important model performances in different aspects. In this study, we employed Back-Propagation (BP) artificial neural network and Classification and Regression Tree (CART) model for model performance comparison in landslide susceptibility prediction. We evaluated model performance from three aspects: Data sensitivity, prediction accuracy, and model stability. The Caiyuan basin in Fujian Province was taken as the study area and 11 landslide-related factors were selected. Additionally, Beichuan county in Sichuan Province was taken as the verification area and 12 landslide-related factors selected. Firstly, two models were both trained using different amounts of data as input. With increasing data volume, the prediction accuracy of BP artificial neural network increased faster than that of CART model. Specifically, in Caiyuan basin, the prediction accuracy of BP artificial neural network and CART decision tree model increased by 5.22% and 2.11%, respectively, for every additional 10 000 samples. In Beichuan county, the prediction accuracy of these two models increased by 4.88% and 3.40%, respectively. Secondly, 100 sets of training data and validation data generated by random sampling were fed into two models for training. The experimental results show that, for Caiyuan basin, the mean prediction accuracy was 81.60% and 72.97% for BP artificial neural network and CART model, respectively, and the standard deviation was 0.32% and 0.35% for BP and CART, respectively. For Beichuan county, the mean prediction accuracy of two models was 77.45% and 72.61%, respectively, and the standard deviation was 0.47% and 0.61%, respectively. Finally, landslide susceptibility maps were generated based on two models. Compared to real landslide spatial distribution map, the result of BP artificial neural network was more consistent with the actual landslide distributions. In general, our study demonstrates that BP artificial neural network is more sensitive to the increase of data volume and has better model stability and prediction accuracy than CART model. But it is worth noting that the performance of two models is close with small data volume. The study provides a new perspective of model selection for landslide susceptibility analysis. |
| |
| Keywords: | landslide landslide susceptibility analysis artificial neural network decision tree comparison analysis prediction accuracy stability sensitivity to data volume |
|
| 点击此处可从《地球信息科学》浏览原始摘要信息 |
| 点击此处可从《地球信息科学》下载免费的PDF全文 |
|
