基于深度学习的火山灾害场景高分遥感检测方法

Detection of volcanic disaster scene from high-resolution remote sensing image with deep learning

针对现有火山灾害场景高分遥感图像智能检测中地表目标类型多样、样本类标缺失问题,提出一种基于深度学习的火山灾害场景高分遥感检测方法.该方法首先以多示例学习网络(Multi-Instance Learning,MIL)为框架,利用联合金字塔上采样(Joint Pyramid Upsampling,JPU)代替扩张卷积,然后通过原型学习和注意力机制(Attention Mechanism,AM)实现对火山灾害场景特征表示的深度神经网络模型重构,并在xBD数据集上进行测试.实验结果表明,与基准卷积神经网络(Convolutional Neural Network,CNN)、MIL方法和“CNN+”深度学习方法相比,在计算耗时未显著增加的情况下,本文方法能够取得最小的标准差和最高的准确性与检测精度,目视效果好.此外,我们进一步利用本文方法对2022年1月14—15日Hunga Tonga-Hunga Ha’apai(HTHH)火山灾害场景多源、多时序高分遥感图像进行检测,与已有成果表现出较好的一致性.

In the existing intelligent detection of volcanic disaster scenes from high-resolution remote sensing images,the diverse types of ground objects and the missing sample labels in volcanic disaster scenes are of critical importance,a detection method based on deep learning for volcanic disaster scene from high-resolution remote sensing image is proposed in this paper.Firstly,based on multi-instance learning(MIL)framework,we use joint pyramid upsampling(JPU)to replace dilated convolution module.Then,using the prototype learning and attention mechanism to reconstruct deep neural network model of volcanic disaster scene feature representation simultaneously.Finally,the xBD dataset is used to test the constructed model performance.The experimental results show that compared with the convolutional neural network(CNN),MIL method,and“CNN+”deep learning methods,our proposed method can achieve the minimum standard deviation and the highest precision and detection accuracy without significantly increasing computational time,and has good visual effects.In addition,we further use the proposed method to detect the Hunga Tonga-Hunga Ha’apai(HTHH)volcanic disaster scenes from multi-source and multi temporal high-resolution remote sensing images on January 14th and 15th,2022,and the detection results have good consistency with existing researches.