Proper prediction of sugarcane distribution rate based on the stem border features is a difficult task because of the irregular forms and morphology, changes in the environmental conditions, and noise during imaging procedures. The proposed paper is a hybrid machine learning architecture that incorporates Capsule and Graph Neural Networks (GNN), Gaussian Process Regression (GPR) and Extreme Learning Machine (ELM) to simulate structural, spatial, and probabilistic sugarcane stem image modeling. The system was trained on one lakh high-resolution border-segmented images and cross-validation done on a large-scale dataset of these images. The fusion strategy that is suggested will have hierarchical texture coding, graph-based boundary representation, and nonlinear regression refinement to increase predecessence and generalization. Compared to traditional deep learning and regression models, experimental evidence proves a higher degree of predictive stability and accuracy of 99.487. The system is a scalable solution to intelligent agricultural analytics, to which it provides automated grading, yield optimization, and precision farming applications.