Improvement of Lidar data classification algorithm using the machine learning technique

Study of the lidar measurement is very significant in a variety of applications including forest remote sensing. Among them, polarimetric lidar is a relatively new but important active remote sensing tool. This study covers a comprehensive description of the system performance of both the polarimetric lidar and non-polarimetric lidar. Noticeably, relative performances of both lidar systems can be estimated exploiting several classifiers such as artificial neural network, k-NN (k-nearest neighbor) classifier and the discriminant function. In each of these attitudes, the principal aspect is to compare the classification results obtained by different classifiers to obtain improved lidar performance. In this case, utility of polarimetric and non-polarimetric waveform features for classification was tested using a group of randomly selected trees such as pines, elm, blue spruce, maple, choke cherry, and green ash. The k-NN classifier obtained 92% accuracy using non-polarimetric data. However, for k-NN classifier the value of k is provided by the user. Strikingly, same k-NN classifier achieved 96 % accuracy using polarimetric data. Again, artificial neural network classifier achieved 96% classification accuracy using polarimetric lidar data whereas the classification accuracy it received was around 89 % using non-polarimetric lidar data. Most poor performance was received by discriminant analysis. In case of non-polarimetric data, discriminant analysis results in only 59 % efficiency. In contrast, about 75 % classification accuracy was observed using polarimetric data for discriminant analysis. Though the listed classifiers can perform better using polarimetric data than non-polarimetric data, artificial neural network can be employed for better performance.