改进深度交叉网络的遥感信息即时智能推荐方法

Improved real-time intelligent recommendation method of remote sensing information in deep cross network

随着遥感大数据时代的到来,海量遥感数据的主动、即时推送问题成为限制遥感信息智能服务领域发展的瓶颈。针对现有遥感信息推荐模型空间特征表达能力不足、交叉特征表达能力欠缺和无差别对待交叉特征等问题,本文提出一种融合注意力机制的深度交叉空间变换网络(attention deep&cross spatial-transformation network,ADCSTN)。首先,模型使用深度交叉网络提取遥感信息不同关联的交叉特征;然后,基于栅格划分,利用空间变换层将一维空间属性数据转换为二维空间矩阵,充分捕捉遥感信息空间结构特征;最后,通过注意力层对得到的不同关联的交叉特征设置不同权重,增强模型性能,实现遥感信息的主动、即时、智能推送。本文利用STK仿真1584颗智能遥感卫星组成的遥感卫星星座为20°N—40°N、120°E—140°E区域内的舰船提供实时遥感数据,并设置用户兴趣,得到试验数据集。试验结果表明,本文模型的推荐效果较好,相比于传统的四元组模型,F_(1)score提高了37%~54.7%。

With the advent of the era of remote sensing big data,the problem of active and real-time push of massive remote sensing data has become a bottleneck limiting the development of remote sensing information intelligent services.Aiming at the problems of insufficient spatial feature expression ability,insufficient cross feature expression ability and non-discriminatory treatment of cross features in existing remote sensing information recommendation models,this paper proposes an attention deep cross spatial transformation network(ADCSTN)integrating attention mechanism.Firstly,the model uses deep cross network to extract the cross features of different associations of remote sensing information.Then,based on grid division,the model uses the spatial transformation layer to convert the one-dimensional spatial attribute data into two-dimensional spatial matrix,fully capturing the spatial structure characteristics of remote sensing information.Finally,the attention layer sets different weights for the different associated cross features to enhance the performance of the model and realize the active,real-time and intelligent push of remote sensing information.In this paper,the remote sensing satellite constellation composed of 1584 intelligent remote sensing satellites is simulated by STK to provide real-time remote sensing data for ships in the area of 20°N—40°N,120°E—140°E,and set user interests to obtain the experimental data set.The experimental results show that the recommendation effect of the model in this paper is better.Compared with the traditional quad model,the F_(1) score is increased by about 50%.