结合深度神经网络的高分辨遥感影像滑坡检测

Landslide Detection from High-resolution Remote Sensing Image using Deep Neural Network

利用遥感影像快速、高效地检测滑坡区域在抢险救援、风险评估等方面发挥着重要的作用。提出了多尺度特征融合的深度神经网络框架用于高分辨遥感影像滑坡自动检测。首先,该框架利用ResNet网络对遥感影像进行筛选,选出具有滑坡区域的候选影像。接着将含有滑坡区域的候选遥感影像输入到多尺度神经网络中对滑坡区域进行分割,从而精准地定位滑坡位置。利用开源数据集和自制数据集对模型进行了训练,并在都香高速公路沿线正射遥感影像上测试。结果表明,该方法在测试区域的滑坡检测总体精度为88.55%,召回率为85.52%。提出的深度神经网络框架能够提取不同尺度的滑坡区域,削弱了与滑坡无关因素的影响,有效提高了高分辨遥感影像滑坡检测的精度。

Quickly and efficiently detect landslide areas by mean of remote sensing images plays an important role in emergency rescue and risk assessment.In this paper,a multi-scale feature fusion deep neural network framework for automatic detection of landslides in high-resolution remote sensing images is proposed.First,the framework uses the ResNet network to screen remote sensing images and select candidate images with landslide areas.Then the candidate remote sensing image containing the landslide area is brought into the multi-scale neural network to segment the landslide area,so as to accurately locate the landslide location.The model is trained using open source data sets and self-made data sets,and tested on orthophotos along the Duyun-Shangri-La Expressway.The results show that the overall accuracy of this method for landslide detection in the test area is 88.55%,and the recall rate is 85.52%.The deep neural network framework proposed in this paper can extract landslide areas of different scales,weaken the influence of factors unrelated to landslides,and effectively improve the accuracy of high-resolution remote sensing image landslide detection.