Leaf Water Potential in a Mixed Mediterranean Forest from Machine Learning and Unmanned Aerial Vehicle (UAV)-Based Hyperspectral Imaging

Netanel Fishman, Yehuda Yungstein, Assaf Yaakobi, Sophie Obersteiner, Laura Rez, Gabriel Mulero, Yaron Michael, Tamir Klein, David Helman

Leaf water potential (ψleaf) is a key indicator of plant water status, but its measurement is labor-intensive and limited in spatial coverage. While remote sensing has emerged as a useful tool for estimating vegetation water status, ψleaf remains unexplored, particularly in mixed forests. Here, we use spectral indices derived from unmanned aerial vehicle-based hyperspectral imaging and machine learning algorithms to assess ψleaf in a mixed, multi-species Mediterranean forest comprised of five key woody species: Pinus halepensis, Quercus calliprinos, Cupressus sempervirens, Ceratonia siliqua, and Pistacia lentiscus. Hyperspectral images (400–1000 nm) were acquired monthly over one year, concurrent with ψleaf measurements in each species. Twelve spectral indices and thousands of normalized difference spectral index (NDSI) combinations were evaluated. Three machine learning algorithms—random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM)—were used to model ψleaf. We compared the machine learning model results with linear models based on spectral indices and the NDSI. SVM, using species information as a feature, performed the best with a relatively good ψleaf assessment (R2 = 0.53; RMSE = 0.67 MPa; rRMSE = 28%), especially considering the small seasonal variance in ψleaf (±σ = 0.8 MPa). Predictions were best for Cupressus sempervirens (R2 = 0.80) and Pistacia lentiscus (R2 = 0.49), which had the largest ψleaf variances (±σ > 1 MPa). Aggregating data at the plot scale in a ‘general’ model markedly improved the ψleaf model (R2 = 0.79, RMSE = 0.31 MPa; rRMSE = 13%), providing a promising tool for monitoring mixed forest ψleaf. The fact that a non-species-specific, ‘general’ model could predict ψleaf implies that such a model can also be used with coarser resolution satellite data. Our study demonstrates the potential of combining hyperspectral imagery with machine learning for non-invasive ψleaf estimation in mixed forests while highlighting challenges in capturing interspecies variability.