滑坡提取卷积神经网络层深结构与显著性特征

Extraction of deep structure and salient features of convolutional neural network from landslide

为了探究影响深度学习滑坡遥感提取的因素,滑坡提取精度随卷积核最优主分特征、层深结构、特征维数的变化规律,获取滑坡提取最优模型,研究深度学习滑坡提取机制。通过可视化卷积特征图和多维散点图技术揭示滑坡提取的地类混淆情况,进而选择支持向量机对深度学习模型的层深结构和特征图分类结果对比。结果显示,18层的SegNet网络模型在样本数量30000时,F1分数为最高的75.38%,最佳滑坡特征分类卷积层层深位置较浅,高于其他层深卷积达10%。研究进一步挖掘出滑坡语义显著性特征,验证了深度学习方法的有效性并将拓宽其在遥感滑坡领域的应用范围,为构建滑坡深度卷积网络指数模型提供显著特征与层深结构基础。

Convolutional neural networks have gradually become a research hotspot in the field of remote sensing extraction of landslides.The deep learning mechanism and mechanism of remote sensing extraction of landslides are studied,and the spatial distribution of semantic features of landslide perimeter images is revealed through layer-depth feature map visualization and multi-dimensional scatter plot technology.Select SVM to classify and compare the visualization results of layer depth structure and feature map,and discover the variation law of landslide extraction accuracy with the optimal principal feature of convolution kernel,model layer depth structure,and model layer depth feature dimension.The SegNet network model,the 14-layer model Conv1,Conv2 and the 18-layer model Conv1 have high classification accuracy,which is 75.38%higher than the output classification accuracy of the model deconvolution layer,and the accuracy is generally higher than other deep convolution layers Up to 10%.It is found that the positive and negative saliency features of landslides have a higher accuracy contribution to the extraction of landslides,and provide significant features and a deep structure basis for the construction of the landslide deep convolution network index model.