Improvements in the acquisition of three-dimensional(3D)information from the Airborne Laser Scanner(ALS)increase its applications for studying Earth’s surface.The use of ALS data in natural resource inventories is st...Improvements in the acquisition of three-dimensional(3D)information from the Airborne Laser Scanner(ALS)increase its applications for studying Earth’s surface.The use of ALS data in natural resource inventories is still in an experimental stage in central Europe.Here,a survey was completed in Germany,where plot-level features from LANDSAT Thematic Mapper and ALS data were applied.An automated process was developed for forest stratification using orthoimages.A genetic algorithm was applied for variable screening.Variable subsets of different sizes were employed for simultaneous predictions of structural forest attributes using the‘Random Forest’(RF)method.Performance was assessed by leave-one-out cross-validations on bootstrap resample data.Results indicate that the stratification of forest notably improved the results of predictions.The improvements were more obvious for the strata-related attributes.Accuracy was enhanced as the number of selected variables increased.However,parsimonious models are still essentially required for practical applications.The RF errors were slightly greater than those from least squares regression,as the non-parametric methods do not share the same mix of error components as regression.Through the combination of remote sensing and modelling,we conclude that our results are helpful for bridging the gap between regional earth observation and on-the-ground forest structure.展开更多
文摘Improvements in the acquisition of three-dimensional(3D)information from the Airborne Laser Scanner(ALS)increase its applications for studying Earth’s surface.The use of ALS data in natural resource inventories is still in an experimental stage in central Europe.Here,a survey was completed in Germany,where plot-level features from LANDSAT Thematic Mapper and ALS data were applied.An automated process was developed for forest stratification using orthoimages.A genetic algorithm was applied for variable screening.Variable subsets of different sizes were employed for simultaneous predictions of structural forest attributes using the‘Random Forest’(RF)method.Performance was assessed by leave-one-out cross-validations on bootstrap resample data.Results indicate that the stratification of forest notably improved the results of predictions.The improvements were more obvious for the strata-related attributes.Accuracy was enhanced as the number of selected variables increased.However,parsimonious models are still essentially required for practical applications.The RF errors were slightly greater than those from least squares regression,as the non-parametric methods do not share the same mix of error components as regression.Through the combination of remote sensing and modelling,we conclude that our results are helpful for bridging the gap between regional earth observation and on-the-ground forest structure.
