Dry beans classification into classes is a critical operation in the agriculture industry, having direct effects on crop quality regulation, pricing, and supply chain effectiveness. Sorting is error-prone and labor-intensive when performed manually, while automated sorting machinery is susceptible to class imbalance and scalability. The paper introduces a machine learning approach for classifying seven dry bean classes with high accuracy using simple morphological characteristics such as area, perimeter, and shape. For addressing extreme class imbalance the paper employs SMOTEN, a multi-class oversampling method. Eight machine learning methods are tested, ranging from logistic regression to decision trees to ensemble algorithms such as Random Forest and XGBoost and deep learning through neural networks. The paper fills crucial gaps by applying neural networks to tabular agriculture classification and by providing the first successful implementation of SMOTEN for multi-class agricultural datasets. The proposed neural network architecture achieves 96.09% classification accuracy, improving on the existing 91% benchmark, while showing excellent precision and recall values across all classes. As opposed to computationally demanding image-based solutions, the proposed solution operates on tabular data, rendering it economically feasible and deployable at large scales. The proposed method provides an economically viable and scalable solution to smart farming use cases.