Abstract: Wildfires have profound impacts on global ecosystems and climate change. Accurately acquiring the information on burned areas is crucial for fire management, carbon emission assessment, and ecological restoration. Traditional burned area extraction methods, such as the MCD64 product, still face limitations in accuracy and spatiotemporal resolution and thus struggle to meet the requirements for high-accuracy monitoring. Hence, this study proposed an improved dual attention network (mDANet) model based on deep learning for burned area segmentation from MODIS images. The mDANet model was trained and validated on the MTBS dataset. Three bi-temporal difference indices (i.e., VI57, NBR2, and MIRBI) were used as input features to enhance the accuracy in detecting burned area boundaries. The experimental results demonstrate that compared to the MCD64 product, the mDANet model exhibited significant improvements in multiple evaluation metrics, including accuracy, recall, and intersection over union (IoU). Specifically, the mDANet model improved the recall value from 0.75 to 0.89 and the mean IoU from 0.50 to 0.72, verifying its effectiveness in burned area extraction. Further visual analysis confirms that the mDANet model produced more continuous burned area boundaries, eliminating the omission errors observed in the MCD64 approach. However, challenges remain in detecting small-scale burned areas, as higher downsampling rates may lead to the loss of some high-frequency spatial information. Overall, the proposed method provides a viable solution for high-accuracy remote sensing monitoring of burned areas.