Fusing small-footprint waveform LiDAR and hyperspectral data for canopy-level species classification and herbaceous biomass modeling in savanna ecosystems

摘要

The study of ecosystem structure, function, and composition has become increasingly important in order to gain a better understanding of how impacts wrought by natural disturbances, climate, and human activity can alter ecosystem services provided to a population. Research groups at Rochester Institute of Technology and Carnegie Institution for Science are focusing on characterization of savanna ecosystems and are using data from the Carnegie Airborne Observatory (CAO), which integrates advanced imaging spectroscopy and waveform light detection and ranging (wLiDAR) data. This component of the larger ecosystem project has as a goal the fusion of imaging spectroscopy and small-footprint wLiDAR data in order to improve per-species structural parameter estimation towards classication and herbaceous biomass modeling. Waveform LiDAR has proven useful for extracting high vertical resolution structural parameters, while imaging spectroscopy is a well-established tool for species classication and biochemistry assessment. We hypothesize that the two modalities provide complementary information that could improve per-species structural assessment, species classication, and herbaceous biomass modeling when compared to single modality sensing systems. We explored a statistical approach to data fusion at the feature level, which hinged on our ability to reduce structural and spectral data dimensionality to those data features best suited to assessing these complex systems. The species classification approach was based on stepwise discrimination analysis (SDA) and used feature metrics from hyperspectral imagery (HSI) combined with wLiDAR data, which could help nding correlated features, and in turn improve classiers. It was found that fusing data with the SDA did not improve classication signicantly, especially compared to the HSI classication results. The overall classication accuracies were 53% for both original and PCA-based wLiDAR variables, 73% for the original HSI variables, 71% for PCA-based HSI variables, 73% for the original fusion of wLiDAR and HSI data set, and 74% for the PCA-based fusion variables. The kappa coecients achieved with the original and PCA-based wLiDAR variable classications were 0.41 and 0.44, respectively. For both original and PCA-based HSI classications, the kappa coecients were 0.63 and 0.60, respectively and 0.62 and 0.64 for original and PCA-based fusion variable classications, respectively. These results show that HSI was more successful in grouping important information in a smaller number of variables than wLiDAR and thus inclusion of structural information did not signicantly improve the classication. As for herbaceous biomass modeling, the statistical approach used for the fusion of wLiDAR and HSI was forward selection modeling (FSM), which selects signicant independent metrics and models those to measured biomass. The results were measured in R2 and RMSE, which indicate the similar ndings. Waveform LiDAR performed the poorest with an R2 of 0.07 for original wLiDAR variables and 0.12 for PCA-based wLiDAR variables. The respective RMSE were 19.99 and 19.41. For both original and PCA-based HSI variables, the results were better with R2 of 0.32 and 0.27 and RMSE of 17.27 and 17.80, respectively. For the fusion of original and PCA-based data, the results were comparable to HSI, with R2 values of 0.35 and 0.29 and RMSE of 16.88 and 17.59, respectively. These results indicate that small scale wLiDAR may not be able to provide accurate measurement of herbaceous biomass, although other factors could have contributed to the relatively poor results, such as the senescent state of grass by April 2008, the narrow biomass range that was measured, and the low biomass values, i.e., the limited laser-target interactions. We concluded that although fusion did not result in signicant improvements over single modality approaches in those two use cases, there is a need for further investigation during peak growing season.