Abstract
The sustainability of agricultural production is increasingly threatened by nutrient deficiencies that compromise crop health, quality, and yield. Optimal crop management requires early detection and classification of these nutrient deficiencies. Traditional methods employ laboratory testing to predict nutrient deficiency in plant samples. Computer vision models provide an alternative to visual inspection by analysing and predicting the presence of nutrient deficiencies. A hybrid feature-based learning framework, AshGdNutDefAM, is proposed for the automated detection of nutrient deficiencies in the Ash gourd ( Benincasa hispida ) plant variety as part of the study. A rule-based and vein-based feature extraction approach is proposed, encompassing domain-specific botanical knowledge along with conventional features such as colour, texture, and spatial descriptors to provide an exhaustive representation of nutrient deficiency symptoms. Key physiological characteristics, such as the green-to-yellow pixel ratio, extent of interveinal chlorosis, and the dimensions of necrotic margins, that are associated with macronutrient and micronutrient deficiencies, especially magnesium (Mg) and iron (Fe), are captured by rule-based features. The proposed framework extracts 55 discriminative features across five groups: colour, spatial, texture, vein, and rule-based features. These features are evaluated using three machine learning classifiers: Support Vector Machine (SVM), Random Forest, and Gradient Boosting. The dataset collected includes healthy, magnesium and iron deficiency of the Ash gourd plant variety captured from agricultural farms in Mysuru, India. The performance of the proposed feature set is further compared with conventional handcrafted feature extraction techniques, including Gray-Level Co-occurrence Matrix (GLCM), Histogram of Oriented Gradients (HOG), and Local Binary Patterns (LBP). The experimental results show that the SVM classifier trained with the proposed feature set achieves an accuracy of 82% and outperforms the conventional handcrafted feature extraction approaches. The dataset used includes only the Ash gourd plant variety and exhibits geographic and temporal biases, as it primarily covers data from specific regions and periods potentially limiting the model’s effectiveness. The proposed design paves the way for intelligent decision-support systems for precision agriculture and sustainable crop management while providing a scalable, comprehensible approach to nutrient deficiency identification.