Nutrient Diagnosis of Eucalyptus at the Factor-Specific Level Using Machine Learning and Compositional Methods

Area of Science:

• Forestry and Agronomy
• Plant Nutrition
• Computational Ecology

Background:

• Brazil holds 30% of global Eucalyptus resources, relying on fertilization for biomass production.
• Current fertilization practices use state standards post-canopy closure, which may not suit local conditions.
• Myriad interacting growth factors at the local scale necessitate customized nutrient diagnosis for Eucalyptus.

Purpose of the Study:

• To customize nutrient diagnosis for young Eucalyptus trees to a factor-specific level.
• To evaluate the effectiveness of local features versus tissue composition in predicting tree yield.
• To compare state nutrient guidelines with locally derived standards for Eucalyptus cultivation.

Main Methods:

• Collected 1861 observations across eight Eucalyptus clones, 48 soil types, and 148 locations in southern Brazil.
• Utilized a random forest classification model to predict high-yielding specimens (cutoff diameter at breast height = 4.3 cm).
• Employed factor-specific diagnosis using Euclidean distance on centered log-ratio transformed compositions of neighboring trees.

Main Results:

• Random forest model achieved an AUC of 0.78 when incorporating local features, compared to 0.63 with tissue composition alone.
• State guidelines showed excessive levels for Mg, B, Mn, and Fe, and deficient levels for Cu and Zn, versus locally balanced specimens.
• Nutrient interactions were overlooked by concentration ranges, while factor-specific diagnosis identified imbalances effectively at the local scale.

Conclusions:

• Downscaling regional nutrient standards may fail due to unaddressed local factor interactions.
• Factor-specific diagnosis, considering local neighbors, is crucial for accurate Eucalyptus nutrient management.
• Large datasets and stakeholder collaboration are essential for documenting local-scale growth factors.