基于注意力融合的遥感滑坡目标识别

Remote sensing landslide target recognition based on attention fusion

针对传统卷积神经网络遥感滑坡识别方法中存在的模型参数量多、重点区域关注不足、难以捕获长期(全局)依赖关系的问题,提出一种融合改进自注意力和卷积块注意力的遥感滑坡目标识别算法。该算法基于编码器-解码器目标识别框架,为了增强模型对滑坡区域局部特征关注程度,将卷积块注意力机制应用于浅层特征提取,从空间与通道两个维度获取滑坡目标特征关联信息。将改进自注意力机制应用于深层特征提取,使模型能够捕获特征图内和特征图间的全局特征信息,实现滑坡目标与背景区域的有效区分。实验结果表明,该方法的滑坡识别精度为96.81%,像素分割准确率均值达到90.11%。通过与FCN、DeeplabV3+等算法进行对比,该方法在保持模型轻量级的同时,有效提升了滑坡识别的准确率。

Aiming at the problems of traditional remote sensing landslide recognition methods based on convolutional neural network,such as too many model parameters,insufficient attention in areas of interest,and difficulty in capturing long-term(global) dependency,an automatic landslide recognition algorithm of remote sensing images based on improved self-attention and convolution block attention is proposed. The algorithm is based on the encoder-decoder target recognition framework. In order to enhance the model’s attention to local features of landslide areas, the convolution block attention mechanism is applied to the extraction of shallow features,and the landslide target feature association information is obtained from the spatial and channel dimensions. The improved self-attention mechanism is applied to the extraction of deep features,so that the model can capture global feature information within and between feature maps,which effectively distinguishes landslide targets from background areas.Experimental results show that the landslide recognition precision of this method is 96. 81%,and the average accuracy of pixel segmentation is 90.11%. The proposed method can effectively improve the accuracy of landslide identification while keeping the model lightweight in comparison with FCN,DeeplabV3+and other algorithms.