MULTI-CLASS CLASSIFICATION OF AIRBORNE LIDAR USING ADABOOST ALGORITHM

High-resolution digital terrain model (DTM), which is a digital representation of ground surface topography, has become invaluable in the area of geomorphometry. High-resolution DTMs are critical for predicting flooding, monitoring erosion, landslide and tectonic movements, modeling ecosystems, and creating digital city models. Airborne LIDAR has recently become an accurate and inexpensive technology to acquire dense point measurements of topography. Such data can be used to generate a high resolution DTM. The DTM generation needs to identify the terrain points on the bare earth, and to remove non-terrain points associated with vegetations, buildings and other man-made objects above the ground. Researchers have developed a number of different filtering algorithms to generate a DTM automatically from the airborne LIDAR data. However, automating the extraction of bare earth model is still challenging. Existing filtering algorithms encounter significant challenges due to the variety and complexity of objects in urban and semi-urban environments. At present, filtering is done semi-automatically, and such approach often consumes an estimated 60%-80% of the processing time. Thus, an automated and robust filtering algorithm is preferred to extract bare-earth model from the LIDAR data. In this research we implemented Adaptive Boosting (AdaBoost) algorithm raw LIDAR data point to perform binary classification (i.e., bare-earth vs non-bare-earth) and multi-class classification (e.g., building, low vegetation, high vegetation, bare-earth). We run the model for a study area of 1.5 km by 0.8 km with 100,000 points (67,000 points for training and the rest for testing). We achieved 95% accuracy for binary classification (bare earth extraction) and 87.3% accuracy for multi-classification of LIDAR data.