Nitrogen Feeds Boxwood Best as Machine Learning Learns to Read Fertilizer Fingerprints

Boxwood, the evergreen shrub that anchors formal gardens across Europe and North America, has long been a quiet workhorse of ornamental horticulture. Yet behind its dense green foliage lies a surprisingly complex question: which fertilizer actually makes it thrive? A new study published in BMC Plant Biology by Ömer Sarı and Elif Enginsu of the Black Sea Agricultural Research Institute in Türkiye, together with Fisun G. Çelikel of Ondokuz Mayis University, set out to answer that question with an unusual combination of tools — meticulous greenhouse measurements of growth, roots, and nutrients, paired with machine learning classifiers trained to recognize the fingerprints each fertilizer leaves on the plant.

The research team evaluated six different fertilizer applications on Buxus sempervirens, the common boxwood: nitrogen, potassium, vermicompost, beneficial bacteria, mycorrhizal fungi, and sheep wool, an organic amendment that has attracted attention as a slow-release nutrient source. The goal was twofold. First, the researchers wanted to quantify how each treatment affected plant height, width, shoot development, leaf dimensions, root architecture, nutrient concentrations, and chlorophyll content. Second, they wanted to test whether machine learning algorithms could distinguish which fertilizer a plant had received simply by reading its aboveground morphology — a question with real implications for how growers might monitor plant nutrition in the future.

The results were strikingly clear for one treatment. Nitrogen emerged as the most effective fertilizer across nearly every measure of plant performance. Plants receiving nitrogen grew 25 percent taller and 34 percent wider than untreated controls, produced shoots 29 percent longer, and developed leaves 18 percent longer. The number of shoots per plant was also numerically higher under nitrogen, rising by 51 percent, although the overall treatment effect on shoot number did not reach statistical significance. Potassium ranked as the second most effective treatment, matching nitrogen’s 25 percent gain in plant height and delivering a 33 percent increase in plant width, along with a 9 percent improvement in shoot length.

Perhaps the most dramatic effects appeared below ground. Root architecture — the hidden half of plant performance that growers rarely see — responded strongly to nitrogen. Root length increased by 45 percent, root surface area by 81 percent, and root volume by a remarkable 119 percent compared with control plants. The number of root forks, the branching points where roots split and explore the soil, was numerically 38 percent higher under nitrogen, though again the overall treatment effect was not statistically significant. These findings matter because a well-developed root system determines how efficiently a plant captures water and nutrients, and how successfully it establishes after transplanting — a critical factor for a species routinely moved from nursery pots into landscape settings.

Nutrient analysis revealed a more nuanced picture. In roots, treatment effects were significant only for nitrogen and potassium concentrations, suggesting that the other measured elements were buffered against change regardless of what was applied. In leaves, the story was different: magnesium, phosphorus, potassium, iron, and manganese all responded significantly to the fertilizer treatments, while calcium, zinc, nitrogen, and copper showed no significant overall effects. Notably, sheep wool — the organic amendment in the trial — showed its strongest influence on selected nutrient responses, particularly potassium, iron, and manganese, hinting that this unconventional material may have a role in fine-tuning micronutrient nutrition even when its growth effects are modest.

Chlorophyll content, a proxy for the photosynthetic machinery that drives growth, varied according to leaf age and the timing of measurement. Nitrogen produced the highest mean chlorophyll values among young leaves measured in June, and nitrogen and potassium treatments generated the highest chlorophyll responses across measurement periods overall. However, the authors report that the overall treatment effect on chlorophyll was not statistically significant, a reminder that what looks like a clear pattern in a single measurement window may not hold across the full season.

The machine learning component of the study added a genuinely novel dimension. The researchers evaluated five classifiers — including J48, a decision-tree algorithm; Naïve Bayes; and Random Forest — using six aboveground morphological characteristics as input features. J48 delivered the highest performance on a hold-out test set, while Naïve Bayes achieved the best five-fold cross-validation accuracy. To understand which traits drove the models’ decisions, the team applied Random Forest feature-importance analysis and SHAP analysis, a technique that attributes each prediction to individual features. Both approaches converged on the same conclusion: plant size and shoot-development characteristics contributed more strongly to the models’ output than leaf dimensions did. In other words, when an algorithm tried to guess which fertilizer a boxwood had received, it relied chiefly on how big the plant was and how vigorously it had shot, not on the fine details of its leaves.

Yet the machine learning results came with an important caveat. While the classifiers identified treatment-related morphological patterns within the dataset, they showed limited generalization to previously unseen observations. This is a common and honest limitation in agricultural machine learning: models can memorize the structure of the data they were trained on without learning relationships robust enough to transfer to new plants, new seasons, or new growing conditions. The authors are candid about this, framing the machine learning analysis as a complementary tool rather than a replacement for conventional measurement — a way to surface which traits matter most, rather than a standalone diagnostic.

That framing points to the study’s broader significance. Fertilizer effects in boxwood proved to be strongly trait-specific: a treatment that excelled at driving height growth might have little effect on chlorophyll, and an amendment that barely moved growth metrics could still shift nutrient concentrations in meaningful ways. By combining conventional plant and nutrient analyses with interpretable machine learning, the researchers demonstrated a workflow that captures this complexity more completely than either approach alone. For nursery producers, the practical takeaway is straightforward — nitrogen remains the workhorse fertilizer for pushing boxwood growth, with potassium close behind, while organic options like sheep wool may earn their place through nutrient-specific effects rather than raw growth gains.

For the wider field of ornamental plant science, the study offers a template. As machine learning tools become more accessible, the challenge is no longer building models but building honest ones — models whose limitations are measured, whose feature importance is interpretable, and whose predictions are tested against data they have never seen. This boxwood trial, modest in scale but rigorous in design, shows what that discipline looks like in practice, and suggests that the future of fertilizer evaluation may belong to teams willing to pair a measuring tape and a root scanner with an algorithm willing to admit what it does not yet know.

Subject of Research: Effects of chemical and biofertilizer treatments on growth, root architecture, and nutrient dynamics of Buxus sempervirens, analyzed with machine learning

Article Title: Growth, nutrient dynamics, and root architecture responses of Buxus sempervirens to chemical and biofertilizer applications: an integrated experimental and machine learning approach

Article References: Sarı, Ö., Enginsu, E., & Çelikel, F. G. (2026). Growth, nutrient dynamics, and root architecture responses of Buxus sempervirens to chemical and biofertilizer applications: an integrated experimental and machine learning approach. BMC Plant Biology. https://doi.org/10.1186/s12870-026-09942-4

Image Credits: AI Generated

DOI: 10.1186/s12870-026-09942-4

Keywords: Buxus sempervirens, boxwood, nitrogen fertilizer, potassium, vermicompost, mycorrhiza, sheep wool, root architecture, machine learning, plant nutrition, ornamental horticulture, chlorophyll