Mapping coastal habitats is essential to their preservation, but the presence of water hinders seamless data collection over land-water interfaces. Thanks to its dual-wavelength and optical properties, topo-bathymetric lidar can address this task efficiently. Topo-bathymetric lidar waveforms contain relevant information to classify land and water covers automatically but are rarely analysed for both infrared and green wavelengths. The present study introduces a point-based approach for the classification of coastal habitats using bispectral waveforms of topo-bathymetric lidar surveys and machine learning. Spectral features and differential elevations are fed to a random forest algorithm to produce three-dimensional classified point clouds of 17 land and sea covers. The resulting map reaches an overall accuracy of 86%, and 65% of the prediction probabilities are above 0.60. Using this prediction confidence, it is possible to map coastal habitats and eliminate the classification errors due to noise in the data, that generate a clear tendency of the algorithm to overestimate some classes at the expense of some others. By filtering out points with a low prediction confidence (under 0.7), the classification can be highly improved and reach an overall accuracy of 97%.