<p>China has the most landslides in the world. How to quickly identify landslides is of great significance for landslide prevention and rescue and relief. This paper aims to improve the landslide recognition ability and generalization ability to adapt to various environments through the combination of deep learning and multi-source data, and provide decision support for early identification of potential landslide risk areas and disaster management. This paper proposes a deep learning module for extracting frequency domain information, called frequency attention, to help the model further mine high-level semantics and improve the segmentation performance of the model. Based on the two-branch architecture and frequency attention, this paper proposes a landslide recognition model for FANET(Frequency Attention Network). On the Bijie landslide dataset, the multi-source data composed of remote sensing images and geological factors are used for training and validation, respectively. The results show that. In both cases, FANET using multi-source data achieves the best results, and its verification accuracy can reach 96.42%, the mean intersection over union ratio can reach 0.8412, the F1 value can reach 0.9091, and the recall rate can reach 0.9066. In order to verify the influence of the combination of multi-source data on the model, seven sets of different geological factor combination experiments are designed. The results show that the combination of DEM, slope and aspect has the best effect, mIoU reaches 0.8537, F1 score is 0.9173, recall rate is 0.9203, precision rate is 0.9144, and accuracy rate is 0.9669. The results show that proper combination of multi-source data can obtain better landslide effect.</p>