Abstract: The detection of cluster distributed targets in remotely sensed satellite images is a challenging task, as cluster is a common behavior of targets and adhesions between dense distributed targets often exist, which affect the accuracy of object detection seriously. However, the distinct distribution pattern of such cluster distributed targets in frequency domain has never been studied. In this paper, a refinement of FFT-based heatmap with multi-branches network for the detection of cluster distributed targets in the satellite images (termed as HeatNet) is proposed. More specifically, a refining method of the FFT-based heatmaps for different features in frequency domain and an attention-based feature extractor in frequency channel are proposed, to focus the attention and refine the salient regions for the cluster distributed targets. Additionally, as one complete system, a keypoint-based detection is adopted as the basic workflow to tackle with the adhesion, a scale-aware center area is conducted to tackle with the variation of scale, and an orientation discrimination is also utilized to eliminate the specificity of different targets. The effectiveness of our proposed method is validated on two public datasets, and the comparative experimental results with different state-of-the-arts object detection methods have demonstrated the superiority of this proposed method.