Hierarchy features attention network for tiny ship detection from SDGSAT-1 thermal infrared images

Accurate and reliable ship target detection is of great significance for the sustainable development goals of ocean management. With the development of remote sensing technology, satellite imagery provides strong support for space-based tiny ship detection. However, remote sensing images have complex backgrounds, and it is challenging to separate and locate different numbers of small ship targets in different scenarios. The thermal infrared bands can capture the temperature differences between ships and the surrounding marine environment, enabling effective detection. Therefore, this study used the Sustainable Development Goals Science Satellite 1 (SDGSAT-1) thermal infrared spectrometer (TIS) to develop a three-channel infrared small target detection (IRSTD) dataset with pixel-level annotations for all bands (SDG-IRSTD). The dataset contains 329 images from the SDGSAT-1 TIS and 3492 targets. Then a hierarchy features attention network (HFA-Net) for space-based tiny ship detection was proposed. The network generates enhanced feature maps of different scales through the multi-level detail enhancement module (MLDEM), employs a multi-level large kernel attention module (MLLKAM) which integrates the multi-scale mechanism with large kernel attention (LKA) to effectively model long-range dependencies on feature maps with different scales, and finally achieves feature fusion and interaction of different scales through the multi-level feature fusion module (MLFFM). In addition, the HFA-Net model improved intersection over union (IoU) and probability of detection (P_d) by 2.35 % and 3.97 %, respectively, and reduced false alarm rate (F_a) by 3.29 × 10⁻⁶, outperforming the state-of-the-art (SOTA) IRSTD methods. It can achieve target localization while obtaining the overall shape of the ship, providing important support for sustainable marine safety.