ALS for terrain mapping in forest environments: an analysis of lidar filtering algorithms.

Remote sensing enables the recording of accurate geomorphological data with the capability to efficiently cover large areas. However, the presence of vegetation makes the use of remote methods for terrain mapping difficult. LiDAR (Light Detection And Ranging) data collection by means of ALS (Airborne Laser Scanning) can be a solution for forestry projects, as the laser pulses cross the entire forest canopy and reach the soil underneath. In order to obtain an accurate digital terrain model, the ALS data must be processed, so as to determine which returns are at ground level. This process is called filtering or classification.
This paper aims to provide a performance analysis of nine algorithms for ALS data classification. The algorithm performance is reviewed for the case of mountainous terrain, characterised by moderate and steep slopes and forest vegetation of a generally high consistency. Out of the nine algorithms tested, two are commercial ones and the others are free.
Our findings suggest that the Lasground-new algorithm implemented in the LAStools (Rapidlasso) software package provides the most accurate results, with a Root Mean Square Error of elevation values for the study site of 0.34 metres (with over 80 percent of the area having an elevation error of less than 0.20 metres) and an average RMSE for the field plots of 0.66 metres. Reference data for RMSE calculation is a DTM interpolated from the ALS point cloud, as classified by the data provider. Some of the free algorithms tested provide relatively similar results in terms of RMSE (for example, MLS and SMRF have RMSE values of 0.56 metres and 0.60 metres, respectively). The correlation between ground slope and RMSE of elevation values is considered for the eight field surveyed plots, with R2 having a value of 0.89.
Taking into account the difficult test conditions (topographically complex surface with dense canopy cover) we consider ALS data to be a possible solution for collecting geomorphological data for forestry applications, as long as data at a relatively low spatial resolution is sufficient.