EFFICIENT DEEP LEARNING: HYPERPARAMETER TUNING OF DENSENET-CNN USING CUCKOO SEARCH FOR ENHANCED ARECANUT DISEASE IMAGE CLASSIFICATION

Arecanut (Areca catechu), a significant commercial crop in tropical region, is vital to the livelihoods of millions of farmers and contributes significantly to the agricultural economy. However, arecanut farming is challenged by several diseases, including Stem Cracking, Yellow Leaf Disease, Mahali (Koleroga), and Stem Bleeding, which negatively affect crop yield and quality. Conventional disease detection methods primarily depend on visual inspection, making the process labour-intensive, time-consuming, and susceptible to errors. To address these challenges, this study aims to develop a robust system for classifying arecanut diseases using a DenseNet-inspired Convolutional Neural Network (CNN) architecture with bottleneck blocks, optimized via Cuckoo search for hyperparameter tuning. Standard CNNs often struggle with feature redundancy, vanishing gradients, and poor generalization, particularly for minority classes. Our approach mitigates these issues by leveraging DenseNet’s feature reuse and an optimized architecture through hyperparameter tuning that enhances gradient flow while reducing overfitting. Compared to a standard CNN, the proposed model balances complexity and accuracy. While CNN achieves a 93.74% test accuracy with lower complexity, it struggles with minority classes. The base DenseNet-inspired model with growth rate 16 underperforms at 87.87% accuracy due to overfitting, whereas the growth rate 32 variant improves to 92.19%. The hyperparameter-tuned model achieves the highest test accuracy of 95.54%, demonstrating superior generalization. However, this comes at the cost of increased model complexity due to feature concatenation and additional trainable parameters. Despite this, the optimized architecture prevents overfitting while enhancing classification, particularly for difficult cases like Stem Bleeding (F1-score 0.84 vs. CNN’s 0.51). This balance makes it the most effective model for disease detection.