This paper presents a method for estimating Manning's roughness coefficients using point cloud data to improve the accuracy of two-dimensional flood inundation simulations. Flood inundation simulations typically use coarse roughness coefficients based on land use classifications, which cannot capture detailed terrain characteristics. Point cloud data were utilized to capture fine-scale variations in terrain elevation and to reflect these variations in the estimation of roughness coefficients. Three approaches for estimating roughness coefficients were examined: (1) the Strickler formula with experimentally determined parameters, (2) the Geomorphometric Statistical Roughness Assessment (GSRA) method, which evaluates terrain characteristics using statistical analysis, and (3) a deep learning-based method trained on hydraulic experiment data. We applied these methods to the Chikuma River basin in Nagano Prefecture, Japan (study area: 9.27 km × 12.6 km, approximately 117 km²), where significant flood damage occurred during Typhoon in 2019. Flood simulations were conducted using roughness coefficients calculated by each of the three methods, as well as using traditional land-use-based roughness classifications. The performance of each approach was evaluated by comparing simulated inundation areas with actual flood extent observed during the 2019 event. The evaluation used the IoU (Intersection over Union) metric to measure how well each method reproduced the observed flood extent. To further enhance simulation accuracy, we also explored methods to address limitations of grid-based modeling, including the detection of elevated structures using point cloud clustering techniques. The proposed method demonstrates that detailed point cloud analysis can provide more accurate roughness coefficient distributions, leading to improved flood inundation predictions.