[Objective] The health status and ecological functions of forest ecosystems were studied in karst areas through remote sensing monitoring of forest volume in order to provide theoretical basis for carbon sink monitoring and assessment, as well as forest management and decision-making in the region. [Methods] Sentinel-2A images and sample plot survey data were acquired for typical karst mountainous areas. Three machine learning models, including random forest (RF), K-nearest neighbor (KNN), and back propagation (BP) neural network, were combined to conduct a study on forest volume inversion under mountain slope conditions. [Results] ① The performance of single-band reflectance, vegetation index, and texture features varied under different topographic constraints, and the optimal subsets of models established were different. There were differences in the establishment of forest volume estimation models under different site conditions. ② For forest volume estimation in the karst mountainous area, RF had the strongest robustness and adaptability compared with KNN and BP. For gentle slope, inclined slope, and steep slope conditions, the accuracy of RF reached 80.1%, 79.0% and 80.5%, respectively. [Conclusion] Karst mountainous areas have strong spatial heterogeneity, and the modeling independent variables involved in the remote sensing estimation of storage volume are not the same under different slope site conditions. Categorizing slope conditions can refine the remote sensing estimation model of complex scenes and improve the accuracy of remote sensing estimation of forest volume.