Examining the utility of a hybrid approach for predicting forest structural attributes using hyperspectral data

Abstract

Mapping the structure of commercial forests using hyperspectral remote sensing techniques produces highly accurate and very fine ecological details that are vital for sustainable forest management and planning. However, the implementation of traditional hyperspectral modelling approaches has been hindered by issues of multicollinearity and data dimensionality. Thus, this study aimed to examine the utility of a hybrid partial least squares regression and random forest (PLSR-RF) algorithm in accurately predicting four structural attributes (basal area, diameter at breast height, tree height and site index) amongst varieties of commercial eucalyptus and acacia using airborne AISA Eagle hyperspectral imagery (2.4 m spatial resolution). Additionally, two traditional hyperspectral modelling methodologies, PLSR and RF, were utilised alongside the hybrid PLSR-RF method. The best model obtained for this study was produced for Acacia mearnsii using the hybrid PLSR-RF algorithm (R ² = 0.92 and RMSE = 0.02). The results of this study indicate that the hybrid approach is a robust and promising model for the estimation and mapping of forest structure for varieties of eucalyptus and acacia within a commercial plantation. In addition, the hybrid model predominantly outperformed the traditional modelling approaches. The hybrid approach successfully navigates the issues of multicollinearity and data dimensionality associated with hyperspectral remote sensing through the combination of PLSR and RF. The generation of in-depth structural information provided by hyperspectral data supports more effective and timely forest management and planning practices.