Comparative Deployment Strategies for Curcumin Concentration Prediction using Hyperspectral Imaging on AWS

Abstract

This study presents a comparative evaluation of Amazon Web Services (AWS)-based deployment strategies for a deep learning model designed to predict curcumin concentration in turmeric (Curcuma longa) rhizomes using hyperspectral imaging (HSI). The proposed hybrid model was trained on features labeled with high-performance liquid chromatography (HPLC) and integrated both spectral and spatial feature extraction. Four distinct deployment pipelines were examined: (i) Amazon EC2-based manual deployment, (ii) fully managed SageMaker inference endpoints, (iii) AWS Step Functions for orchestrated workflows, and (iv) Kubernetes-based deployment via Amazon Elastic Kubernetes Service (EKS). Each approach was systematically assessed across dimensions of scalability, cost efficiency, automation, monitoring, and operational effort. The comparative analysis reveals that SageMaker offers the most balanced solution, combining ease of setup, monitoring, and auto-scaling, while Step Functions excel in modular orchestration. EC2 and EKS provide higher control but at the expense of operational overhead. Findings highlight AWS SageMaker as the optimal strategy for research transitioning toward production-scale agricultural machine learning applications, particularly for real-time curcumin estimation in hyperspectral imaging workflows.