Ship Detection in Dual-Polarization SAR Imagery Based on YOLO and Late Fusion

Abstract

With the recent development of satellites and microsatellite constellations, higher temporal resolution satellite imagery and larger numbers of satellites have become available. As the availability of satellite data increases, it is necessary to develop ship detection models that consider specific data characteristics, such as polarization, spatial resolution, and frequency bands. Among these, to leverage the distinct features of polarization, we employed a late fusion approach to merge the ship detection results from Sentinel-1 dual-polarization data. To evaluate the effectiveness of the fusion model, we built four single training models using two polarizations (VV, VH) and two colormaps (gray, parula), as well as six multimodal models with late fusion. As a result of comparing the accuracy of the single model and the fusion model, we found that the accuracy of the fusion model consisting of 1) VH gray colormap and VV parula colormap, and 2) VH parula colormap and VV parula colormap is higher than that of the single model (based on intersection over union (IoU) thresholds of 0.4 and 0.5). Each fusion model achieved a relative accuracy improvement of at least 1.5% and up to 6.5% compared to the single model with the highest accuracy among the two. The significance of this study is that the late fusion was applied using both polarization data and colormap information simultaneously. These results suggest that the fusion model can detect ships more accurately and that colormaps, which have been underexplored in SAR research, can be a factor in improving accuracy.