目标检测算法在滑坡识别中的应用

Application of the object detection algorithm in landslide identification

基于目视解译的滑坡数据集选取两大类(6个)目标检测算法构建滑坡自动识别模型,以四川省阿坝州作为研究区进行滑坡自动识别.基于高分辨率卫星影像构建包含3 120个样本的滑坡样本数据集;选取4种一阶段检测算法YOLOv5(s、 m、 l、 x)与两种二阶段检测算法Faster R-CNN(VGG16和ResNet-50)分别构建滑坡自动识别模型;为探究样本数量对模型识别精度的影响,将样本数据集总数分为1 000、 2 000和3 000,通过滑坡测试样本对识别结果进行评价.结果表明,基于目标检测的两类滑坡识别模型中,一阶段YOLOv5模型比二阶段Faster R-CNN模型更适用于滑坡识别;样本数对滑坡识别模型的性能具有一定影响.在较少样本的情况下,选择YOLOv5s模型能够获得较高的识别精度,随着样本数的增加,使用YOLOv5m模型可以获得更好的滑坡识别效果.

Two major categoriesof target detection algorithms(six in number)were selected based on the landslide database of visual interpretation to construct a corresponding automatic landslide identification model,and Aba Prefecture,Sichuan Province was taken as the study area to conduct a research on automatic landslide identification.A landslide dataset used high-resolution satellite imagery containing 3120 samples.Four one-stage detection algorithms,i.e.YOLOv5(s,m,l,and x),as well as two two-stage detection algorithms,Faster R-CNN(VGG16 and ResNet-50)were employed to build corresponding landslide recognition models.In order to investigate the influence of the sample number on the model recognition accuracy,the total number of sample datasets was divided into 1000,2000,and 3000.The recognition results were evaluated by landslide test samples,which showed that,of the two categories of object detection models for landslide recognition,the one-stage YOLOv5 models were more suitable than the two-stage Faster R-CNN models.The number of samples influenced the performance of the landslide recognition model.In the case of fewer samples,the YOLOv5s model was selected to obtain a higher recognition accuracy,while with the increase in the number of samples the YOLOv5m model could be used to obtain better landslide recognition results.