In this study, we introduce an innovative frequency-area-weighted sampling method to address spatial and temporal biases in flood inventory creation. Focusing on Thailand’s Nan River Basin, we integrated 13 flood conditioning factors and developed a point-based inventory that includes 3000 flood and 3000 non-flood samples, proportionally allocated on the basis of flood recurrence intervals and spatial distribution. We evaluated four machine learning models—artificial neural network, support vector machine, K-nearest neighbors, and random forest (RF) models—to assess their performance in flood susceptibility mapping (FSM). Among these, the RF model demonstrated the highest predictive capability, achieving an area under the curve (AUC) of 0.979 for the test set and an AUC of 0.984 for the verification set. The resulting susceptibility map identified 10.64% of the study area as “very high” risk, providing critical insights for prioritizing flood mitigation efforts. This work advances FSM methodology by effectively bridging the temporal flood frequency and spatial heterogeneity in inventory design, offering a robust framework for data-driven flood risk management in vulnerable regions.