Abstract
Accurate and timely recognition of chilli leaf disorders is essential for reducing crop loss and supporting efficient disease management in practical agricultural environments. However, multi-class chilli leaf disease classification remains challenging because visually similar symptoms can arise from fungal infection, pest damage, nutritional deficiency, and normal leaf appearance, while class imbalance and limited interpretability often reduce the practical reliability of deep learning systems. To address these issues, this study proposes a unified multi-variant lightweight convolutional neural network (CNN) framework for five-class chilli leaf disease classification, comprising
cercospora
,
healthy
,
mites_and_trips
,
nutritional
, and
powdery mildew
. Four custom CNN variants, namely Conv2DNet, MaxPoolNet, GAPNet, and Dense256Net, were evaluated under identical conditions to provide a controlled architectural comparison. The framework combines imbalance-aware training through data augmentation, class weighting, label smoothing, and validation Macro-F1-guided checkpointing; comprehensive reliability assessment through class-wise metrics, confusion matrices, ROC and precision–recall curves, confidence analysis, Grad-CAM, cross-validation, statistical testing, ablation analysis, and external validation; and deployment-oriented comparison using computational complexity and inference measures. A focused head ablation isolates the contribution of the Dense(256) projection, while the baseline evaluation is extended to EfficientNetV2-B0 and ConvNeXt-Tiny in addition to the existing standard and lightweight models. Dense256Net achieved a test accuracy of 0.928, a Macro-F1 score of 0.918, and a weighted F1-score of 0.925. Five-fold cross-validation confirmed stable performance with a mean accuracy of (0.930 ± 0.003) and Macro-F1 score of (0.920 ± 0.004). The model retained a favorable profile with 2.12M parameters, 0.28G FLOPs, an 8.1 MB model size, and 4.8 ms/image GPU inference time. External validation on a publicly available independent chilli leaf dataset achieved 0.904 accuracy and 0.891 Macro-F1 without retraining. Overall, Dense256Net provides the strongest performance–efficiency balance among the proposed variants and remains competitive with the evaluated baselines.