Forest biometric characterization through remote sensing applications

Abstract

The general objective of this thesis was to step forward in the application of aerial 3D-data in the forest characterization context. To meet this goal, the thesis focused on four cutting-edge research topics related to the forestry applications of 3D data collected by airborne laser scanning (ALS) and digital aerial photogrammetry (DAP). Four common algorithms were deeply investigated to filter ALS ground points. The results showed that performing exhaustive filter calibration is not mandatory to derive accurate digital terrain models (DTM), that the applications of software defaults can derive accurate DTM as well, and that filter calibration has a significant but low practical improvement on the prediction of forest attributes using area-based approach (ABA). The application of the high-flexible Johnson’s SB probability density function (PDF) was adapted to the ALS data and compared with the Weibull PDF to estimate diameter distributions in two forest stands, an eucalyptus stand and a radiate pine stand. Johnsons SB was highly sensitive to the prediction of the inputs used to fit the parameters, reasons why this function was just slightly better than Weibull. The ALS data from five different forest sites were used to compare three common modeling approaches used to estimate growing stock volume, ordinary least squares (OLS), random forest (RF), and k-nearest neighbor (kNN). The estimation was more accurate with OLS and RF. The kNNbased models had the worst prediction accuracy and may result in overfitting. The point clouds derived from ALS and DAP presented comparable results when it comes to detect and estimate individual tree volumes in eucalyptus plantations. This result benefits the DAP since it is an inexpensive approach to collect 3D forest data, especially when associated with unmanned aerial vehicles