Machine Learning-Based Classification of Nitrogen Nutritional Status in Common Bean Across Phenological Stages Using RGB Image Texture

Nitrogen (N) fertilization is essential for common bean (Phaseolus vulgaris L.) yield, and image-based texture analysis offers a fast, non-destructive alternative to conventional nutritional diagnosis. While previous studies have already shown that Red-Green-Blue (RGB) texture attributes can discriminate N levels in common bean, this study addresses two aspects that remain underexplored: the joint effect of phenological stage and image block size on classifier performance. A greenhouse dataset was built using cultivar Madrepérola (BRS FC104) grown in 30 pots across five N doses (0, 50, 100, 150, and 200% of the recommended rate, corresponding to 0, 0.58, 1.17, 1.75, and 2.33 g N per pot; six replicates each). Images were acquired at stages V3, V4, and R1, and cropped into blocks of four sizes (10 × 10, 40 × 40, 60 × 20, and 80 × 80 pixels), yielding 6000 image blocks per dose (30,000 in total). To prevent data leakage, training and test sets were strictly separated at the pot level, ensuring that all image blocks from a given plant were assigned exclusively to one set. Texture features were extracted via the gray-level co-occurrence matrix (GLCM) and classified using Support Vector Machine(SVM), Random Forest, K-Nearest Neighbors (KNN), and Convolutional Neural Network (CNN). Performance varied with stage and block size: CNN was the most stable classifier (mean accuracy 71.8%), while KNN and SVM performed best at later stages and larger blocks (peaks of 80.8% and 83.2%, respectively). These results support GLCM texture descriptors as a promising tool for N-status classification, warranting validation on independent plants under field conditions.