A machine learning approach to model leaf area index in Eucalyptus plantations using high-resolution satellite imagery and airborne laser scanner data

As a forest structural parameter, leaf area index (LAI) is crucial for efficient intensive plantation management. Leaf area is responsible for the energy absorption needed for photosynthetic production and transpiration, both affecting growth. Currently, LAI can be estimated either by remote-sensing methods or ground-based methods. However, unlike ground-based methods, remote estimation provides a cost-effective and ecologically significant advance. The aim of our study was to evaluate whether machine learning algorithms can be used to quantify LAI, using either optical remote sensing or LiDAR metrics in Eucalyptus dunnii and Eucalyptus grandis stands. First, empirical relationships between LAI and remote sensing data using LiDAR metrics and multispectral high-resolution satellite metrics, were assessed. Selected variables for LAI estimation were: forest canopy cover, laser penetration index, canopy relief ratio (from among the LiDAR data), the green normalized difference vegetation index, and normalized difference vegetation index (from among spectral vegetation indices). We compared the accuracy of three machine learning algorithms: artificial neural networks (ANN), random forest (RF) and support vector regression (SVR). The coefficient of determination ranged from 0.60, for ANN, to 0.84, for SVR. The SVR regression methods showed the best performance in terms of overall model accuracy and RMSE (0.60). The results show that the remote sensing data applied through machine learning algorithms provide an effective method to estimate LAI in eucalypt plantations. The methodology proposed is directly applicable for operational forest planning at the landscape level.