Abstract

Smartphone-based proximal sensing has emerged as a promising low-cost approach for pasture monitoring. A critical component of this methodology is accurate leaf segmentation, as it directly affects the reliability of subsequent image-based analyses. Despite advances in computer vision, the role of preprocessing strategies in segmentation performance remains insufficiently explored, particularly under resource-constrained conditions. This study presents a systematic comparative evaluation of three preprocessing pipelines based on HSV and CIELab color spaces for the segmentation of Urochloa grass leaves (Urochloa hybrid Mavuno and Urochloa decumbens) using smartphone imagery acquired field conditions. The pipelines were assessed using a multi-criteria framework, including the Fisher Discriminant Ratio (FDR), Intersection over Union (IoU), Overlap Error (OE), Structural Similarity Index (SSIM), and Edge Preservation Index (EPI), complemented by discordance map analysis. The results demonstrate that preprocessing design significantly influences segmentation stability, boundary preservation, and robustness to illumination variability. Pipelines based on HSV channels showed high sensitivity to shadows and non-uniform lighting, leading to reduced segmentation consistency. In contrast, the CIELab-based pipeline relying on the a* channel achieved superior performance, with higher discriminative capacity, improved edge preservation, and lower computational cost. These findings highlight that carefully designed classical preprocessing strategies remain highly effective for low-cost, real-time applications, even in the absence of computationally intensive models. This work establishes a robust segmentation foundation for future integration with advanced analytical methods, including machine learning approaches, and supports the development of scalable smartphone-based tools for pasture monitoring.

IPC Classification

Keywords