Feature Model Engineering in Machine Learning for Pipeline Variability Management and Plant Disease Recognition in Brazilian Flora

Abstract

This work presents the engineering and empirical evaluation of a feature model for variability management in Artificial Intelligence (AI) pipelines applied to plant disease recognition in Brazilian flora. The proposed artifact, named PAIES (Pipeline Artificial Intelligence Engineering System) feature model, systematically organizes the configurable dimensions of an AI pipeline, data source, preprocessing, model training, evaluation and deployment, allowing the derivation of distinct configurations oriented by functional and non-functional quality attributes. Two configurations were derived and evaluated on the PlantVillage dataset [4]: Pipeline A (performance-oriented, using SwinTransformer and cloud deployment) and Pipeline B (efficiency-oriented, using MobileNet and edge deployment). The results indicate that Pipeline A achieves F1-Score of 0.9931 and accuracy of 0.9923, while Pipeline B achieves F1-Score of 0.9593 and accuracy of 0.9607, representing a loss of only ∼3.4 percentage points. However, the non-functional difference is substantial: Pipeline A presents latency 3.2× higher, model size 6.4× larger and energy consumption 4.7× greater. The results allow rejecting H0 in the context of this empirical study, indicating that the feature model may enable the prediction of significant non-functional trade-offs before model training.