ChangeHabitats2

Classification method

The processing method used in CH2 uses field data collected during field work and remote sensing data (full-waveform aerial laser scanning – FWF-ALS) together to extract relevant habitat features. The processing methods include transforming already pre-processed 3D point-clouds into multiple raster layers derived from 3D point attributes, then extract pixels corresponding to reference/training data from all the rasters to train a machine learning classifier (e.g. Random Forest). Some part of reference data is set aside to serve as an independent validation dataset. After a machine-learning model is trained, its accuracy is estimated by comparing its prediction against a validation dataset. Accuracies and error estimates are reported in a standard classification report, and a confusion matrix is included to support detailed understanding of classification performance. The model is then used to predict vegetation / features map of the entire study site. A number of different classification scenarios can be automatically processed, each resulting in a classification and map extracting different ecologically relevant aspects of the site (e.g. land cover, species, structure, artificial objects and disturbances).

The classification workflow is largely automated using the Vegetation Classification Studio (VCS) software framework, allowing for fast and semi-automated execution of the whole processing chain – from FWF-ALS derived rasters to final high-resolution maps and reports for all classification scenarios. The output maps are stored in standard GeoTIFF files in multiple resolutions and formats, depending on end-user’s needs. The maps can readily be used for further (geo-)processing using regular GIS analysis tools.