Detectors based on convolutional neural networks (CNN) commonly employ label assignment to distinguish positive and negative samples during training. However, existing label assignment strategies overlook the diverse characteristics of objects in remote sensing images(RSI), such as arbitrary directions, large aspect ratios, and varying scales, which leads to inadequate and low-quality sample issues. To tackle these challenges, we propose a novel distance-sensitive label assignment (DSLA) strategy to effectively select both adequate and high-quality positive samples. Specifically, we design an elliptical region sampling (ERS) strategy to carefully screen candidate positive samples by utilizing elliptical regions, thereby mitigating background interference that hampers the model’s performance. Furthermore, we propose a distance-controlled compensation loss (DC-Loss) to further enhance the effectiveness of ERS by reducing the impact of low-quality samples. Extensive experiments are conducted on two challenging datasets for rotated object detection, namely DIOR-R and HRSC2016, validate the superiority of our proposed method.