Wildfires have been one of the most damaging natural disasters in history, severely damaging the global ecological environment, socio-economics, and public safety. To predict and generate wildfire susceptibility maps to visualize the probability of wildfire occurrence intuitively, we propose a Clustering-based Resampling with the Cost-sensitive Boosting (CRCBoost) method. Emphasizing the highly cluster-able feature vectors in wildfire imbalance data overlooked by previous studies, CRCBoost adopts a clustering-based resampling strategy in each iteration of boosting and sets various misclassification costs for positive and negative samples to be more sensitive for identifying the fire points. Due to the frequent wildfires in two contract counties (Los Angeles and Ventura) in California, USA, they are set as the study area. Data regarding historical wildfires, geographical, meteorological, and vegetation features are collected for the period spanning 2010 to 2020. A recursive feature elimination algorithm is used for filtering out irrelevant features. CRCBoost is more suitable for short-term wildfire occurrence prediction and its performance is demonstrated to be better than other popular machine learning methods. Generating wildfire susceptibility maps and calculating feature importance are helpful to identify areas of high wildfire susceptibility and main drivers. Moreover, the findings can further assist in wildfire prevention, mitigation, and response.