Differentiation of ecosystems in the outback of Australia using airborne LiDAR

Abstract

Monitoring ecosystems plays a key role in facing climate change impacts. A better understanding of ecosystem functioning is needed to take appropriate actions toward their conservation. Because of their endemism, their vital services to local populations, and their fragility caused by human pressures and climate change, ecosystems of the outback of Australia are of major importance. Remote sensing is widely used for large-scale ecosystem monitoring however, drylands remote sensing faces unique challenges not typically encountered in other regions. Their high heterogeneity and the soil background reflectance are just a few of the difficulties encountered in these lands. The innovative LiDAR technology has the advantage of detecting the three-dimensional structure of the vegetation and could potentially overcome the uncertainty generated by optical imaging in arid and semi-arid areas. The general goal of this work is to study the potential use of high resolution airborne LiDAR data to classify different ecosystems found in the outback of South Australia. In order to characterize the ecosystems, various structural components were calculated from LiDAR point clouds. The relevance of these metrics in the discrimination of our ecosystems was assessed through a PCA coupled with an analysis of their descriptive statistics. Three classification models were build (a hierarchical clustering, a decision trees and a LDA) with different numbers of input variables. Their performance was compared with the accuracy related to the confusion matrix. The two-variable, top of canopy height and number of trees, models offer the best compromise between parsimony and accuracy. The LDA is the best predictor with an accuracy of 84%. However, the decision tree, whose overall accuracy is only 1% less, is easier to interpret and to relate to the ecological reality. Then, in order to evaluate the ability of the models to be upscaled, discriminant models were tested on larger plots, but the results were not satisfactory. This is due to the number of trees, used as input variable, being dependent on the size of the plot. The use of GEDI data was also explored to assess the potential of models to be extrapolated to the global scale. However, the comparison of airborne and spaceborne LiDAR metrics revealed a significant difference between the two datasets. The results confirm the hypothesis that metrics derived from high resolution airborne LiDAR are capable of discriminating some Australian ecosystems. Nonetheless, this study is a first approach and further research are required to improve the large-scale characterization of drylands ecosystems.Monitoring ecosystems plays a key role in facing climate change impacts. A better understanding of ecosystem functioning is needed to take appropriate actions toward their conservation. Because of their endemism, their vital services to local populations, and their fragility caused by human pressures and climate change, ecosystems of the outback of Australia are of major importance. Remote sensing is widely used for large-scale ecosystem monitoring however, drylands remote sensing faces unique challenges not typically encountered in other regions. Their high heterogeneity and the soil background reflectance are just a few of the difficulties encountered in these lands. The innovative LiDAR technology has the advantage of detecting the three-dimensional structure of the vegetation and could potentially overcome the uncertainty generated by optical imaging in arid and semi-arid areas. The general goal of this work is to study the potential use of high resolution airborne LiDAR data to classify different ecosystems found in the outback of South Australia. In order to characterize the ecosystems, various structural components were calculated from LiDAR point clouds. The relevance of these metrics in the discrimination of our ecosystems was assessed through a PCA coupled with an analysis of their descriptive statistics. Three classification models were build (a hierarchical clustering, a decision trees and a LDA) with different numbers of input variables. Their performance was compared with the accuracy related to the confusion matrix. The two-variable, top of canopy height and number of trees, models offer the best compromise between parsimony and accuracy. The LDA is the best predictor with an accuracy of 84%. However, the decision tree, whose overall accuracy is only 1% less, is easier to interpret and to relate to the ecological reality. Then, in order to evaluate the ability of the models to be upscaled, discriminant models were tested on larger plots, but the results were not satisfactory. This is due to the number of trees, used as input variable, being dependent on the size of the plot. The use of GEDI data was also explored to assess the potential of models to be extrapolated to the global scale. However, the comparison of airborne and spaceborne LiDAR metrics revealed a significant difference between the two datasets. The results confirm the hypothesis that metrics derived from high resolution airborne LiDAR are capable of discriminating some Australian ecosystems. Nonetheless, this study is a first approach and further research are required to improve the large-scale characterization of drylands ecosystems.