Adaptive Multi-Scale Pulse Coupled Neural Network With Gradient-Based Optimization For Enhanced Neem Leaf Disease Classification: A Deep Learning Framework For Precision Agriculture

Abstract

Neem (Azadirachta indica) trees are economically vital for sustainable agriculture and pharmaceutical industries, but foliar diseases significantly compromise productivity and bioactive compound quality. Traditional manual disease detection methods are labor-intensive, subjective, and error-prone, necessitating automated solutions for large-scale monitoring and early intervention.

This study presents a novel Adaptive Multi-Scale Pulse Coupled Neural Network with Gradient-Based Parameter Optimization (AMS-PCNN-GPO) integrated with multi-head attention mechanisms for precise neem leaf disease classification and early detection. The proposed framework introduces five key technological innovations: (1) gradient-based PCNN parameter optimization using hybrid Particle Swarm Optimization (PSO), (2) multi-head attention integration with PCNN architecture for disease-specific feature enhancement, (3) Dynamic Threshold Adaptation Mechanism (DTAM) based on local image statistics, (4) hierarchical multi-scale feature fusion processing images at 1×, 2×, and 4× resolutions, and (5) comprehensive loss function combining classification accuracy, attention consistency, and feature similarity optimization. The system was evaluated on a comprehensive dataset of 2,400 high-resolution neem leaf images (512×512 pixels) across six major disease categories: Alternaria Leaf Spot, Bacterial Blight, Colletotrichum Leaf Spot, Damping Off, Leaf Web Blight, and Powdery Mildew.

Extensive experimental validation demonstrates superior performance compared to state-of-the-art methods. The AMS-PCNN-GPO framework achieves 94.7% classification accuracy, 93.8% precision, 94.2% recall, and 94.0% F1-score, representing significant improvements of 7.3%, 6.4%, 6.8%, and 7.6% respectively over baseline PCNN approaches. The integration of PSO-optimized parameters reduces computational complexity by 35% while maintaining high accuracy. Statistical validation through 10-fold stratified cross-validation confirms robustness (p < 0.001 for all performance comparisons). The system demonstrates 85% accuracy in critical early-stage disease detection and enables 70% reduction in manual inspection costs. The AMS-PCNN-GPO framework establishes new performance benchmarks for automated plant disease classification, combining neuromorphic processing advantages with modern deep learning optimization. The system's processing capacity of 1,200 images per hour enables real-time monitoring for large-scale precision agriculture applications, contributing significantly to sustainable farming practices, food security, and agricultural automation. This research demonstrates the successful integration of bio-inspired computing with artificial intelligence for practical agricultural solutions.