基于EfficientNet的滑坡遥感图像识别方法——以贵州省毕节市为例

Landslide Remote Sensing Image Recognition Based on EfficientNet:Taking Bijie City,Gui⁃zhou Province as an Example

近年来,随着工程建设的快速发展,工程活动改变了边坡原始地质条件,导致滑坡灾害频繁发生,严重威胁人民的生命财产安全。因此,深入研究滑坡的快速、精确识别方法对于防灾减灾具有重要意义。本文提出一种基于EfficientNet高效网络提取滑坡深度特征的潜在滑坡识别方法,该方法通过寻找一组最优的复合系数从深度、宽度、分辨率三个维度对神经网络进行扩展,自适应地优化网络结构,并引入带动量的梯度下降算法(Stochastic Gradient Descent Momentum,SGDM)作为网络学习的优化器,充分考虑历史梯度的影响,在参数更新过程中不断调整当前梯度值,从而相应地调整参数的更新幅度,改善神经网络的学习效果,提取滑坡体的深层次特征。实验结果表明,EfficientNet模型在测试集上的平均准确度达到92.78%,可以高效准确地实时提取滑坡信息,对灾后的快速反应有指导意义。

In recent years,with the rapid development of engineering construction,engineering activities have changed the original geological conditions of slopes,resulting in frequent landslide disasters,which seri-ously threaten people's life and property safety.Therefore,it is of great significance to study the rapid and ac-curate identification method of landslides for disaster prevention and reduction.In this paper,a potential land-slide recognition method is proposed based on EfficientNet to realize the extraction of landslide depth fea-tures.The method extends the neural network from three dimensions of depth,width,and resolution by search-ing for a set of optimal composite coefficients,and adaptively optimizes the network structure.The Stochastic Gradient Descent Momentum(SGDM)is introduced as the optimizer of network learning,which fully consid-ers the influence of historical gradient.And the current gradient value is constantly adjusted during the param-eter updating process,so as to adjust the parameter updating amplitude accordingly,improve the learning ef-fect of neural networks and extract the deep features of the slope.The experimental results show that the aver-age accuracy of the EfficientNet model on the test set reaches 92.78%,which can efficiently and accurately ex-tract landslide information in real-time and provides guiding information for the rapid response after the disaster.