Classification of Land-Water Continuum Habitats Using Exclusively Airborne Topobathymetric Lidar Green Waveforms and Infrared Intensity Point Clouds

International audienceCoastal areas host highly valuable ecosystems that are increasingly exposed to the threats of global and local changes. Monitoring their evolution at a high temporal and spatial scale is therefore crucial and mostly possible through remote sensing. This article demonstrates the relevance of topobathymetric lidar data for coastal and estuarine habitat mapping by classifying bispectral data to produce 3D maps of 21 land and sea covers at very high resolution. Green lidar full waveforms are processed to retrieve tailored features corresponding to the signature of those habitats. These features, along with infrared intensities and elevations, are used as predictors for random forest classifications, and their respective contribution to the accuracy of the results is assessed. We find that green waveform features, infrared intensities, and elevations are complimentary and yield the best classification results when used in combination. With this configuration, a classification accuracy of 90.5% is achieved for the segmentation of our dual-wavelength lidar dataset. Eventually, we produce an original mapping of a coastal site under the form of a point cloud, paving the way for 3D classification and management of land and sea covers.Coastal areas host highly valuable ecosystems that are increasingly exposed to the threats of global and local changes. Monitoring their evolution at a high temporal and spatial scale is therefore crucial and mostly possible through remote sensing. This article demonstrates the relevance of topobathymetric lidar data for coastal and estuarine habitat mapping by classifying bispectral data to produce 3D maps of 21 land and sea covers at very high resolution. Green lidar full waveforms are processed to retrieve tailored features corresponding to the signature of those habitats. These features, along with infrared intensities and elevations, are used as predictors for random forest classifications, and their respective contribution to the accuracy of the results is assessed. We find that green waveform features, infrared intensities, and elevations are complimentary and yield the best classification results when used in combination. With this configuration, a classification accuracy of 90.5% is achieved for the segmentation of our dual-wavelength lidar dataset. Eventually, we produce an original mapping of a coastal site under the form of a point cloud, paving the way for 3D classification and management of land and sea covers