SEQUENTIAL DENSE NEURAL NETWORK FOR TOMATO LEAF DISEASE DETECTION USING BATCH NORMALIZATION AND DUAL REGULARIZATION

Abstract

Tomato leaf diseases are difficult to distinguish reliably when symptoms overlap or images are acquired under inconsistent conditions. This study presents a compact Sequential Dense Neural Network, or SDNN, for classifying nine tomato diseases and healthy foliage. The experimental record contains 8,043 images resized to 224 × 224 pixels. The network combines three convolutional blocks with batch normalization, max pooling, dense classification layers, dropout, and L2 regularization. It was trained with Adam and categorical cross-entropy for a maximum of 100 epochs. The recorded run reports 96.50% training accuracy, 96.67% validation accuracy, 96.71% precision, 96.68% recall, and a 96.69% F1-score. The accompanying figures also report an 18 MB model footprint and a training-time value of 2.5. These outcomes suggest a favorable accuracy–efficiency balance, but internal checks identify unresolved evidence issues. The displayed confusion matrix contains more observations than the stated dataset, all class-wise ROC curves show AUC = 1.00 despite non-zero classification errors, the training-time unit is missing, and the stated flattened dense design appears inconsistent with an 18 MB float32 model. The results are therefore retained and interpreted as reported outputs rather than independently reproduced measurements. A publication-grade rerun should use a frozen split, saved predictions, a fully specified averaging rule, an unregularized baseline, component ablations, and hardware-normalized timing. The model remains a promising screening architecture, but its quantitative claims require reproducible verification before they can support field deployment.