Abstract

The estimation of leaf water content (LWC) is essential for understanding the physiological status of forest species and assessing fire hazards. This study explores the potential of unmanned aerial vehicle UAV-acquired multispectral imagery for LWC estimation in Pinus radiata and Eucalyptus globulus. These are two commercially valuable timber species, commonly planted in monospecific stands in mediterranean forests and widely used in similar plantations worldwide, often prone to wildfires. Leaf samples were collected from individual trees located in two plots in the Valpara & iacute;so region of Chile. Leaves were brought to the laboratory and LWC estimated through gravimetric methods. Five types of models were trained to estimate LWC from UAV-acquired imagery, including classical and machine learning techniques: Linear Regression, Mixed Linear Regression, Random Forest (RF), Support Vector Regression (SVR), and k-Nearest Neighbors (k-NN). These models were trained to predict LWC based on the spectral reflectance data obtained from the UAV imagery. Results from both LWC and spectral models exhibit distinctive physiological traits of each species, with P. radiata showing more conservative water strategies and higher water retention compared to E. globulus. The best-performing model for E. globulus was a mixed linear regression, achieving an R-2 of 0.69, an RMSE of 9.38, and a bias of -1.44. For P. radiata, the best model was a simple linear regression with an R-2 of 0.33, an RMSE of 10.29, and a bias of -0.68. The mixed linear regression model of E. globulus also provided more consistent predictions and better spatial mapping of LWC distribution.