基于ANN与GIS技术的区域岩溶塌陷稳定性预测——以桂林西城区为例

ANN- and GIS-Based Regional Prediction of Cover-Collapse Probability: A Case Study in West Part of Guilin City

岩溶地面塌陷是岩溶区常见的一种地质灾害,塌陷区域预测是进行国土规划、资源开发与灾害防治的必要工作.由于岩溶塌陷的影响因素众多且相互作用,发展过程复杂,加之各评价因子的数值获取困难,致使长期以来塌陷区域定量预测成为一个难以解决的课题.现行的区域预测模型不能描述塌陷形成模式的非线性特征,也难以克服评价因子权重确定过程中人为经验因素的影响.神经网络技术的自学习、自适应与高度非线性映射特点显示了其在塌陷区域预测领域中应用的前景.根据研究区内地面塌陷空间聚集分布的特征,提出了不同因子组合条件下塌陷发生可能性的定量化方法,结合选定的评价因子类别确定了神经网络预测模型的结构,利用312个塌陷点样本中的292个进行网络训练,余下的20个样本的校验结果表明该模型具有较高的可信度.运用GIS技术将研究区进行评价单元划分,并获取各评价因子的取值,输入到训练好的网络中进行预测.将各单元的输出值进行归并处理后得到研究区岩溶塌陷的稳定级分区图.

Covercollapse is one of frequent geological hazards in karst zone. Predicting the probability of covercollapse occurrence is a requisite task for territorial planning, resource exploiting and hazard harnessing. Regional quantitative prediction of covercollapse probability has become very intricate problem for the following reasons: (1) interactions between many influencing factors; (2) a sophisticated developing process; (3) difficulty associated with the value acquisition of factors. Some recent prediction models can not display the nonlinear characteristics of collapse development pattern, nor can they eliminate the impact of empiricism during the course of weights allocation. Three major characteristics of artifical neural network (ANN) technology, i.e. selflearning, selfadapting and nonlinear mapping, indicate a powerful application potential in collapse prediction field. This paper reports the methodology of developing an ANN model to predict covercollapse occurrence probability. An approach was established to measure the relative probability of the collapse corresponding to certain factor combinations, and some impact factors were specified. Consequently, the structure of ANN prediction model was created. 292 stochastic collapse samples from the sample aggregate, of which the size was 312, were used to train the ANN model. The testing results of the 20 remaining samples show that this model has a good precision. Evaluation grids division and their value acquisition of every factor were accomplished with the aid of GIS software tools. The unstable probability of each grid was calculated through the trained ANN model, which enables us to delineate the different stable zones in the study area.