The regression tree-based model developed in this activity was designed to predict real numerical outcomes, such as the mineralization of soil organic matter, and to identify relationships among the dataset variables representing site conditions. The model iteratively partitions the feature space using axis-parallel hyperplanes, generating intervals as homogeneous as possible with respect to the target variable.
The model was trained on real experimental data from the long-term experimental platform of Tetto Frati (UNITO), which includes nine experimental plots sampled over 29 years. The considered variables included:
Meteorological: 7 temperature-related variables (Tmax, Tmin, average temperature, thermal excursion, degree-days at 0, 5, and 10°C), 4 variables related to the water balance (water input, ET0, water balance, dry period), and 3 consolidated climatic indices calculated over 5 temporal scales.
Pedological: 3 soil characteristics.
The following figure shows an example of the data used in the experimental tests.
In a detailed zoom on the regression tree, each node displays the site condition (dataset variable) and the value that best separates or divides the instances to predict soil organic matter mineralization. This process aims to reduce data impurity.
Specifically, a node represents a decision point: it tests the value of a variable, such as a temperature or soil characteristic, and determines the path to follow based on that value. The node condition splits the data into subsets that are more homogeneous concerning the target variable, like the degree of mineralization.
This splitting continues iteratively, creating a tree-like structure, where each branch leads to further decision nodes or terminal leaf nodes. The leaf nodes contain the final predicted value, representing the model’s prediction for a specific combination of conditions.
This interpretability allows users to understand precisely how different site conditions influence the prediction, making the model transparent and easy to analyze visually and statistically.
The experimental tests conducted on the agricultural data provided by the partners showed a Mean Absolute Percentage Error (MAPE) of 22.1%, confirming the model’s ability to provide accurate and interpretable estimates for the analyzed problem.
This translation is precise and maintains the technical meaning and tone suitable for research reports and scientific documentation.
