International Journal of Science, Engineering and TechnologyAuthors: Parag Yadav, Assistant Professor Raj Kumar, Shantanu Yadav, Tanya Rana, Yash Singh Kathayat
Abstract: Accurate, rapid, and non-destructive assessment of sugarcane quality is a critical requirement for optimizing harvesting schedules, refinery throughput, and fair payment systems in the global sugar industry. Conventional wet chemistry methods, while accurate, are time-consuming, reagent-intensive, and unsuitable for real-time field deployment. Near-Infrared (NIR) spectroscopy has emerged as a powerful analytical technique for non-destructive quality measurement; however, existing machine learning approaches applied to NIR spectral data predominantly rely on shallow models or architectures that treat samples in isolation, thereby neglecting the rich relational and interaction structure latent within spectral populations. In this paper, we propose a novel end-to-end deep learning framework, designated EFI-GNN-NIR, that integrates 1D Convolutional Neural Network (CNN)-based spectral feature extraction with an Explicit Feature Interaction-Aware Graph Neural Network (EFI-GNN) for simultaneous prediction of Brix percentage, Pol percentage, and purity index of sugarcane samples. The proposed framework constructs a population-level spectral similarity graph wherein each node represents an individual NIR sample and edges encode cosine similarity relationships between learned spectral embeddings. An explicit feature interaction module, employing bilinear cross-network operations within the graph message-passing paradigm, enables the model to capture higher-order cross-wavelength dependencies that conventional models overlook. The framework is evaluated on a benchmark NIR sugarcane dataset augmented with agronomic metadata, demonstrating superior performance with R2=0.978R^2 = 0.978 R2=0.978, RMSE=0.142RMSE = 0.142 RMSE=0.142 Brix, and RPD=6.84RPD = 6.84 RPD=6.84 for Brix prediction, outperforming state-of-the-art baselines including Partial Least Squares Regression (PLS-R), Support Vector Regression (SVR), Random Forest (RF), standard CNN, and vanilla GNN architectures. Multi-task learning across three quality parameters yields consistent improvements over single-task counterparts. This work provides a significant contribution toward precision agriculture, smart sugarcane farming, and real-time quality monitoring pipelines deployable at factory intake points.
DOI: http://doi.org/
