Multi-pathway neural network architecture with feedback-based validation learning for robust plant disease classification

Abstract

Automated plant disease classification from leaf images demands models that jointly achieve high accuracy and efficient training convergence. Standard deep learning approaches process images through a single feature pathway, limiting their ability to capture diverse visual manifestations such as color changes, texture patterns, and spatial distributions. This paper introduces DMCNNA-FBVL, a framework integrating two complementary innovations: (1) a Deep Multi-Component Neural Network Architecture (DMCNNA) employing three specialized pathways processing color-space statistics, texture descriptors, and raw image features before fusing them through SoftMax-weighted aggregation; and (2) Feedback-Based Validation Learning (FBVL), a training strategy that periodically blends validation-set gradients into weight updates to accelerate convergence. Experiments on the New Plant Diseases Dataset (87,000 images, 38 classes) show that DMCNNA-FBVL achieves 98.7% accuracy, 98.8% precision, 98.6% recall, and 98.7% F1-score, outperforming ResNet-50 by 3.2 percentage points ( 𝑝 < 0 . 0 0 1 ). The primary reported metrics are computed exclusively on an independent 10% hold-out test set, whereas the separate 10% validation partition is used during training for FBVL gradient blending and does not contribute to final test evaluation. Five-fold cross-validation confirms stability (98.7% ± 0.15%). Ablation experiments confirm additive gains, while FBVL reduces wall-clock training time by 15% through faster convergence.