To improve the accuracy and efficiency of flood inundation mapping, a flood monitoring framework based on a lightweight deep learning model and multi-source remote sensing data fusion was proposed. Taking the 2022 “22·6” catastrophic flood in the Beijiang River Basin as a case study, based on Sentinel-2 multispectral optical images and GF-3 synthetic aperture radar images, the flood inundation range was extracted using threshold segmentation method and M-UNet deep learning model, and the disaster situation was analyzed in combination with land use types. The results indicate that the overall accuracy of water extraction using the threshold segmentation method exceeded 95%. The overall accuracy of the M-UNet model is 98.20%, with an average intersection to union ratio of 88.60% and a Kappa coefficient of 92.45%. Compared with the UNet model and DeeplabV3+model, it exhibits excellent flood remote sensing recognition ability. The flood inundation range extracted by integrating multi-source remote sensing data is basically consistent with the post disaster field survey results, among which Yingde, Qingcheng, Qingxin, and Fogang in Qingyuan are the most severely affected, and cultivated land is the main type of land use affected by the disaster.