基于卷积神经网络的高分遥感影像露天采矿场识别

Opencast Mining Area Recognition in High-Resolution Remote Sensing Images Using Convolutional Neural Networks

矿山环境监测常用遥感影像,研究卷积神经网络(convolutional neural network,CNN)对高分遥感影像露天采矿场识别有利于提高监测效率.针对因训练数据量小而导致CNN对露天采矿场识别精度不高的问题,采用了3种迁移学习方案对CNN的预训练模型进行训练与对比分析,发现冻结CNN预训练模型底层参数并微调高层参数的迁移学习方法训练效果最佳,在验证数据上的生产者精度与用户精度均超过87%.实验结果表明,本方法训练的CNN能提升高分遥感影像中露天采矿场的识别效率,可作为遥感解译露天采矿场中的辅助手段.

The application of remote sensing is a commonly used approach to environmental monitoring in mine areas. The research on convolutional neural network (CNN) for recognition of opencast mining area in high-resolution remote sensing image could help to improve monitoring efficiency. This paper focused on the problem of low classification accuracy of opencast mining area based on CNN due to insufficient trained datasets, the experiment is designed with three types of transfer learning methods and tested in different pre-trained CNN models. The analysis shows that by contrast, fixed lower layers’parameters in pre-trained CNN models and fine-tune higher layers’parameters is the optimal training method, it achieved over 87% in both producer’s accuracy and user’s accuracy. This experimental results indicate that opencast mining area could be effectively recognized in high-resolution imagery based on this training method, therefore, the CNN which trained by this method can be used as an aid in remote sensing interpretation of opencast mining area.