基于GCN模型的毕节金沙石场乡集镇岩溶塌陷易发性评价

Assessment of karst collapse susceptibility in Shichang Township, Jinsha County, Bijie City based on GCN model

摘要

【目的】为解决当前区域岩溶塌陷易发性研究在空间捕捉和非线性拟合方面的不足，提出基于图卷积神经网络(Graph Convolutional Networks)的建模框架。【方法】结合岩溶发育分形特征定量化评估岩溶发育程度，并以贵州省毕节市石场乡集镇为研究区，选取16个孕灾因子建立GCN塌陷易发性评价模型，同时与支持向量机、随机森林模型进行对比的方法研究了区域岩溶塌陷易发性。【结果】结果显示：同为岩溶区不同地层对岩溶塌陷影响存在差异，将复杂岩溶形态转化为分形维数可实现岩溶发育程度定量化评估，有效规避主观判断偏差；三种评价模型均能改善非线性拟合不足问题，极高/高风险区走向与研究区岩溶槽谷走向大致一致且模型存在相似性与差异性，其中GCN模型的准确率、灵敏度、特异度、精确率、AUC均显著高于其他两种模型。GCN模型的极高易发区呈现连续条带状分布，与研究区岩溶槽谷走向契合，重合率达92%并且GCN模型在极高易发区集中了33个灾害点(占比75%),且该区域面积比仅18.18%,体现出对高易发性区灾害点的强聚集性识别能力。【结论】GCN模型结合岩溶发育分形特征，可精准识别研究区中部、西部的极高易发区，其空间分布与岩溶强发育带完全契合，比传统单一指标评价更能捕捉塌陷风险的空间异质性。研究成果增进了对毕节市金沙县石场乡岩溶发育程度的认识，同时建立了一套较为适合的岩溶塌陷易发性评价体系，GCN模型在区域岩溶塌陷易发性评价中效果更为明显。

Abstract

[Objective] To address the deficiencies in spatial feature capturing and nonlinear fitting in current regional karst collapse susceptibility studies, a modeling framework based on graph convolutional networks(GCN) is proposed. [Methods] The fractal characteristics of karst development were integrated to quantitatively evaluate the degree of karst development. Shichang Township in Bijie City, Guizhou Province, was selected as the study area. 16 disaster-inducing factors were selected to establish a GCN-based collapse susceptibility evaluation model, and the regional karst collapse susceptibility was investigated by comparing the GCN model with support vector machine(SVM) and random forest(RF) models. [Results] The result showed that different strata in the same karst area had varying impacts on karst collapse. Converting complex karst morphologies into fractal dimensions enabled the quantitative assessment of karst development degree, effectively avoiding subjective judgment bias. All three evaluation models mitigated the issue of insufficient nonlinear fitting, and the orientation of very high/high susceptibility zones was roughly consistent with the trend of karst troughs in the study area, and the models exhibited both similarities and differences. Among them, the accuracy, sensitivity, specificity, precision, and AUC of the GCN model were significantly higher than those of the other two models. The very high susceptibility zones identified by the GCN model displayed a continuous strip-like distribution, which aligned with the orientation of the karst troughs in the study area, with an overlap rate of 92%. Moreover, the GCN model concentrated 33 hazard points(accounting for 75%) in these very high susceptibility zones, while the area of these zones only accounted for 18.18% of the total area, demonstrating the model's strong capability to identify the clustering of hazard points in high susceptibility zones. [Conclusion] By integrating the fractal characteristics of karst development, the GCN model accurately identifies the very high susceptibility zones in the central and western parts of the study area, and the spatial distribution is fully consistent with the strong karst development zones. Compared with traditional single-indicator evaluation, the model captures the spatial heterogeneity of collapse risks more effectively. The findings enhance the understanding of the karst development degree in Shichang Township, Jinsha County, Bijie City. Additionally, a relatively suitable evaluation system for karst collapse susceptibility is developed. The GCN model demonstrates more prominent performance in regional karst collapse susceptibility evaluation.

关键词

岩溶发育程度 / 洼地密度 / 分形特征 / 机器学习 / 易发性评价 / 地质灾害 / 影响因素 / 图卷积神经网络

Key words

karst development degree / depression density / fractal characteristics / machine learning / susceptibility evaluation / geological hazards / influencing factors / graph convolutional networks

基金资助

国家重点研发计划“深部岩溶塌陷风险精准防控技术装备研发”(2022YFC300330)