Spaceborne remote sensing image object detection based on mixture of experts model

As an important application direction in the field of spaceborne remote sensing, object detection is of great significance in disaster emergency response, resource management, environmental monitoring and other fields. We take ship targets as an example to study object detection in spaceborne remote sensing images. Spaceborne remote sensing covers sea areas in different regions worldwide, and these sea areas have significant differences in geomorphic features. Existing ship target detection methods mostly use a unified structure to handle different scenarios, making it difficult to adapt to changes in ground object features in different regions and with different resolutions. To address this, we introduce the Mixture of Experts (MoE) model into the task of spaceborne remote sensing image object detection, and propose a spaceborne remote sensing image object detection model based on the Mixture of Experts model. It adaptively selects corresponding expert processing for the differential features of different geographical regions to achieve more accurate ship target detection in different sea areas. The model includes two groups of experts: one group is guided by spatial resolution information to capture differences in ground object scales under different resolutions; the other group uses latitude and longitude information for guidance to adapt to the distribution characteristics of ground objects in different geographical regions. Through the collaborative work of the two groups of experts, the model can more accurately perceive and locate ship targets in images, significantly improving detection performance. In addition, we construct a dataset containing multi-resolution information and latitude and longitude information-the Geographic In-formation Ship Detection Dataset (GISD) to provide data support for the research. We conduct experimental verification on the GISD dataset and the public dataset LEVIR-ship. The average precision AP50 of the proposed method reaches 80.6% and 85.7% respectively, which is 3.8% and 2.0% higher than that of the best-performing models on the two datasets.