For object detection on Remote Sensing Images (RSI), numerous methods based on deep convolutional neural networks have been developed by researchers(CNN) and and re-markable achievements have been made in detection performance and efficiency. Current CNN-based methods usually require a large number of annotated samples for training. However, labeling RSI is time-consuming, making it difficult to obtain large-scale annotated training samples. In this paper, we introduce a transfer learning-based method for few-shot object detection on RSI. In our method, only a few annotated samples are required for unseen classes. More specifically, our model adopts a two-stage fine-tuning scheme and contains two modules: a multi-scale self-attention module and a copy-paste with diminishing edge transparency module. Our design enables the model to learn transferable knowledge from seen classes and generalizes well to unseen classes. Experiments on two benchmark datasets demonstrate the effectiveness of our proposed method in few-shot object detection for RSI.