Data annotation quality in smart farming industry

Abstract

ABSTRACT Data has become paramount in modern agriculture industry, empowering precision farming practices, optimizing resource utilization, facilitating predictive analytics, and driving automation. However, the quality of data influences the usefulness of smart farming systems. Poor data quality specifically in annotations include mislabeling, inaccuracy, incompleteness, irrelevance, inconsistency, duplication, and overlap. These limitations often emerge from specific constraints, or the small scale of the problem/limited number of stakeholders, making it challenging to overcome the obstacles. This paper presents a comprehensive framework to address data annotation quality challenges. Beginning with raw, non-curated data, we integrate three strategic methodologies: active learning, enhanced annotations, and transfer learning to elevate the data quality resulting in a curated data model that achieves superior data quality refining decision-support. Leveraging a curated set of KPI, we demonstrate that integrating data quality measures leads to enhanced accuracy, reliability, and performance. Transfer learning is the most promising approach, demonstrating superior performance.