Performance analysis of deep feature extraction, feature fusion and feature selection with machine learning techniques in classification of chickpea seeds

Abstract Chickpeas are an essential food source rich in carbohydrates and proteins. However, distinguishing among different chickpea varieties remains challenging due to their high morphological similarity. Most recent studies rely on handcrafted or traditional feature-based methods, which are time-consuming, sometimes destructive, and often lack generalization capability. To overcome these limitations, we present a non-destructive computer vision and machine learning framework for efficient chickpea seed classification. Deep features were extracted using convolutional neural networks, fused to capture complementary representations, and reduced through the Tree–Seed Algorithm (TSA), a metaheuristic optimization method for feature selection. The selected features were classified using various machine learning algorithms, achieving a maximum accuracy of 95.6%. TSA reduced the feature dimensionality by approximately 60%, significantly decreasing training time while preserving high accuracy. Compared with existing studies reporting accuracies between 83% and 94%, the proposed approach improves classification performance by up to 12%. To the best of our knowledge, this is the first study to integrate deep feature fusion with TSA-based feature selection for chickpea seed classification. The results demonstrate a robust, efficient, and non-destructive alternative to conventional approaches.