Remote sensing image recommendation based on spatial-temporal embedding topic model

摘要

ABS T R A C T Through research and analysis of the existing remote sensing image sharing and distribution systems, remote sensing image recommendation mode can be divided into subscription recommendation and active recommendation. The first mode provides data query retrieval and subscription distribution services. However, retrieval and subscription services are based on query and subscription keywords, which are problematic or insufficiently active for users. Moreover, these processes cannot discover the latent requirements of a user. Therefore, how to recommend remote sensing images to users accurately and actively is a challenging problem. Research on the active remote sensing image recommendation is rare. The typical method is a space-time periodic task model (STPT), which realizes personalized remote sensing image recommendation based on simulation user log records. However, STPT is not accurate enough because it uses the minimum bounding rectangle as the filter condition of spatial feature and considers that the user's acquisition of images is periodic, so the data that match the periodic rules is more likely to be returned, resulting in a low recall rate. Additionally, it is less efficient for large-scale image recommendation tasks because it takes a long time to calculate the time-period using Fourier transform method. In this study, we propose a spatial-temporal embedding topic (STET) model to solve the recommendation problem of remote sensing images. This model processes the spatial, temporal, and content information of remote sensing images and constructs a topic model, thereby fully applying the continuity characteristics of space and time and improving the training efficient of the recommendation model. Compared with state-of-the-art models, the results based on large scale real-world datasets show that our model not only significantly improves the recall by more than 10%, the normalized discounted cumulative gain by more than 10% when K is 100 with the precision remaining above 97%, but it also greatly reduces the training time.