基于三维卷积与残差结构的沟谷泥石流危险度评价

Risk assessment of debris flow in valleys based on stereoscopic convolution and residual structure

沟谷的泥石流危险度评价是泥石流防治工作中基础且重要的一环,针对该问题,以怒江州为例,提出了一个结合立体卷积与残差结构,能同时对数字高程模型(Digital Elevation Model,DEM)数据与多光谱数据进行特征学习的神经网络模型.以整沟为研究对象,将模型在历史泥石流灾害沟谷的数据上训练后,根据相似度对沟谷的泥石流危险度进行评分,并绘制了怒江州的泥石流危险度评价图.在所有214条沟谷中,高风险沟谷共114条,中风险沟谷共40条,低风险沟谷共60条.实验结果表明,该模型能在沟谷泥石流分类任务上达到最高80%的正确率、88%的召回率以及0.81的Kappa系数.此外,在使用更少训练数据的实验以及对比各个不同模型的实验中,所设计的模型均表现优异.模型给出的危险度与历史灾害记录和实地考察结果基本相符.

The risk evaluation of the valley debris flow is a basic and important part in preventing debris flow.To make the evaluation,we combine the stereoscopic convolution and the residual structure,and propose a neural network model for feature learning on DEM and multi-spectral data simultaneously.This article takes Nujiang Prefecture as an example and investigates the susceptibility of the entire-valley debris flow.After training on the data of historical debris flow hazard valley,the network can give a reasonable evaluation of the hazard of the debris flow based on the debris flow similarities,and then the risk assessment map of debris flow disaster in Nujiang Prefecture is drawn.Among all 214 valleys,there are 114 high risk valleys,40 medium risk valleys,and 60 low risk valleys.The experimental results show that the model can achieve the highest accuracy rate of 80%,the recall rate of 88%and Kappa coefficient of 0.81 on the task of classification of valley debris flow.In addition,in experiments that use less training data and experiments that compare different models,the designed models perform better.The degree of risk given by the model is consistent with the disaster facts and the result of field investigations.