New Data Processing Approach for Vegetation Classification using Multiwavelength Airborne Polarimetric Lidar

Polarimetric measurement has been proved to be significant in a variety of applications including forest remote sensing. Polarimetric full waveform lidar is a relatively new but important active remote sensing tool. This instrument has opened up the possibilities for extracting the full waveform data along with the polarimetric information. So far, a number of research works have been carried out using commercial non-polarimetric lidar data for tree species classification, at either dominant species level or individual tree level. Many of these classification approaches used a combination of features including tree height, stand width and crown shape. However, the research work presented in this thesis has addressed the problem using an approach based on polarimetric lidar waveforms collected from remotely located trees. This study presents research results with a number of illustrating examples depicting the potential of polarimetric lidar data in capturing additional information associated with the polarimetric reflection. Moreover, this lidar instrument would be an answer to obtaining spectral reflectance of tree crown at more than a single wavelength.;In fact, only relatively little research has been done with full-waveform polarimetric lidar. So one of the aims of this thesis is to develop a comparative study between the polarimetric lidar and the non-polarimetric lidar based on their ability on vegetation species classification. The key idea of this scheme is to separate the trees on the basis of the dominant plant categories. Other strategy follows more challenging approach in tree separation: identification of trees at individual level. Both strategies utilize similar data from same type of trees. In order to deal with the data collection problems, lidar data from five different tree species were digitally sampled in the field. These include ponderosa pine, austrian pine, blue spruce, green ash and maple. The time history of the light interaction with the target for a complete backscattering cycle was recorded as lidar waveforms. Both polarized and simulated non-polarized data were calibrated using a calibration standard to remove the discrepancies of lidar returns. Afterwards, whole data were randomly split in two data sets, namely training set and test set. Training data set was used to train and generalize the classifiers whereas test data set was used to verify the classification result. To accomplish this, full-waveform lidar data were used without any further processing. In another approach, same lidar signals were preprocessed applying power spectral density (PSD) analysis and resultant dimensional redundancy were removed using principal component analysis (PCA). Relative performances of both lidar systems were estimated with the exploitation of this data at the input of several standard classifiers such as artificial neural network (ANN), k-NN classifier and discriminant function. In the first scheme, the polarimetric lidar was able to separate the deciduous trees from the coniferous trees with a higher accuracy than that of non-polarimetric lidar. On the other hand, lidar data was also successfully capitalized to yield observable partitions among the individual trees. These studies support the fact that polarimetric reflectance distribution along the lidar profile could be attributed to the higher tree separability at both levels. Classification results, along with the exploratory data analysis demonstrated that polarimetric lidar data is more accurate than the traditional non-polarimetric lidar data in terms of classification accuracy.