Smart agriculture increasingly relies on machine learning (ML) systems in settings where labeled data are costly to obtain, become available progressively over time, and may reflect changing operating conditions. These characteristics make static train-once workflows difficult to sustain in practical applications, especially when data become available in a season-wise manner under non-stationary conditions, and call for iterative model refinement. This paper presents AL-MLOps, a reproducible workflow that combines Active Learning, model optimization, and experiment tracking within an MLOps workflow. The framework is evaluated on a plant health classification task using an augmented PlantVillage dataset in a simulated incremental data-availability setting. In this setting, the results indicate that the proposed workflow maintains predictive performance comparable to full-dataset training while requiring approximately 87% fewer labeled samples. These findings suggest that combining Active Learning with MLOps can support more data-efficient model refinement in smart agriculture.
AL-MLOps: An Active Learning–Driven Framework for Smart Agriculture
Aqasizade, Hossein
;Antonini, Mattia;Vecchio, Massimo;Antonelli, Fabio
2026-01-01
Abstract
Smart agriculture increasingly relies on machine learning (ML) systems in settings where labeled data are costly to obtain, become available progressively over time, and may reflect changing operating conditions. These characteristics make static train-once workflows difficult to sustain in practical applications, especially when data become available in a season-wise manner under non-stationary conditions, and call for iterative model refinement. This paper presents AL-MLOps, a reproducible workflow that combines Active Learning, model optimization, and experiment tracking within an MLOps workflow. The framework is evaluated on a plant health classification task using an augmented PlantVillage dataset in a simulated incremental data-availability setting. In this setting, the results indicate that the proposed workflow maintains predictive performance comparable to full-dataset training while requiring approximately 87% fewer labeled samples. These findings suggest that combining Active Learning with MLOps can support more data-efficient model refinement in smart agriculture.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
