【摘要】 针对滑坡易发性识别中卷积神经网络(CNN)预测效果和模型的利用效率被低估的问题，本文提出一种基于层自适应卷积神经网络(LA-CNN)的滑坡易发性评估方法，使用层自适应卷积神经网络对卷积神经网络的结构进行优化，并引入空间金字塔池化(SPP)以处理输入全连接层的图像。以门源回族自治县为例，进行了滑坡易发性计算，并使用随机森林(RF)和逻辑回归(LR)两种传统机器学习方法和卷积神经网络(CNN)用于对比试验。精度评估结果显示:LA-CNN的ROC曲线下面积(AUC)和模型的准确度(ACC)分别为96.68%和90.35%,AUC值相比于LR、RF和CNN分别提升了3.12%、1.98%和1.14%，准确度相比于LR、RF和CNN提升7.89%、2.63%和0.88%。更多还原

【Abstract】 This paper proposes a landslide susceptibility assessment method based on layer-adaptive convolutional neural network( LACNN) to address the issue of underestimation of the prediction performance and model utilization efficiency of convolutional neural network( CNN) in landslide susceptibility recognition. The structure of the convolutional neural network is optimized using layer-adaptive convolutional neural network,and spatial pyramid pooling( SPP) is introduced to process images input into fully connected layers. Taking Menyuan Hui Autonomous County as an example,landslide susceptibility calculation was conducted,and two traditional machine learning methods,Random Forest( RF) and Logistic Regression( LR),as well as Convolutional Neural Network( CNN),were used for comparative experiments. The accuracy evaluation results show that the area under the ROC curve( AUC) and model accuracy( ACC) of LA-CNN are 96.68% and 90.35%,respectively. The AUC values have increased by 3.12%,1.98%,and 1.14% compared to LR,RF,and CNN,respectively,and the accuracy has increased by 7.89%,2.63%,and 0.88% compared to LR,RF,and CNN.更多还原