# Enhanced paddy leaf disease detection using novel dual metaheuristic loss functions in generative adversarial networks with identity block preservation for thermal image augmentation

## TL;DR

This paper introduces a new GAN framework with dual metaheuristic loss functions to improve thermal image augmentation for paddy leaf disease detection.

## Contribution

A novel dual metaheuristic loss function framework in GANs with identity block preservation for thermal image augmentation.

## Key findings

- The proposed GAN achieves 31.47 dB PSNR and 0.923 SSIM, outperforming existing methods.
- Vision Transformer reaches 97.89% accuracy with the new augmentation, surpassing standard techniques.
- The framework shows robustness under environmental variations and generalizes well across datasets.

## Abstract

This paper presents a novel dual metaheuristic loss function framework integrated within Generative Adversarial Networks (GANs) for enhanced thermal image augmentation, specifically designed to improve paddy leaf disease detection through intelligent data quality enhancement and diversity generation. The proposed methodology revolutionizes traditional GAN training by replacing conventional loss functions with two bio-inspired metaheuristic algorithms: the Chaoborus algorithm, which serves as an innovative generator loss function implementing intelligent missing pixel imputation through phantom midge larvae hunting behavior simulation, and the Australian Crayfish algorithm, which functions as an advanced discriminator loss function optimizing adaptive 8-pixel connectivity through foraging and territorial behavior modeling. The framework incorporates strategically positioned identity blocks to preserve critical thermal signatures during adversarial training, ensuring disease-specific thermal patterns remain intact throughout the image enhancement process while maintaining diagnostic integrity. The proposed dual metaheuristic GAN achieves superior image generation quality with 31.47 ± 0.52 dB Peak Signal-to-Noise Ratio (PSNR) and 0.923 ± 0.008 Structural Similarity Index Measure (SSIM), representing significant improvements over state-of-the-art methods including StyleGAN2 (26.89 dB PSNR), Progressive GAN (27.34 dB PSNR), and BigGAN (28.12 dB PSNR). Disease classification performance evaluation across four distinct neural network architectures (ResNet-50, EfficientNet-B7, Vision Transformer, and DenseNet-201) reveals substantial accuracy improvements, with the Vision Transformer achieving 97.89 ± 0.63% accuracy using the proposed augmentation compared to 83.45 ± 1.76% on original datasets and 87.23 ± 1.54% with standard augmentation techniques. Statistical significance analysis confirms the robustness of improvements with p-values less than 0.001 for all comparative metrics and Cohen’s d effect sizes exceeding 1.2, indicating large practical significance. Rigorous tenfold cross-validation yields consistent performance with 96.85% mean accuracy and low standard deviation (0.674%), while Leave-One-Out Cross-Validation demonstrates minimal bias (< 0.0012) and low variance (< 0.0055). Generalization studies across five different datasets show robust transferability with direct transfer accuracies ranging from 84.12% to 91.45%, improving to 89.67–95.89% with minimal fine-tuning. Environmental robustness evaluation reveals excellent stability under varying temperature (15–35 °C), humidity (40–80%), and temporal conditions, with performance drops limited to 6.07% under extreme conditions. Comprehensive ablation studies validate the synergistic contribution of each framework component, with individual algorithms providing 5.68 dB and 3.76 dB PSNR improvements respectively, while their combination with identity blocks achieves the full 31.47 dB performance. Computational efficiency analysis demonstrates practical viability with 45.2 ± 2.8 ms inference time for image generation and 1.41 × speedup over baseline methods while maintaining 9.8 GB memory efficiency. Real-world field deployment across four geographic locations in India (Punjab, Tamil Nadu, West Bengal, and Odisha) over 2.75 months processing 44,860 images achieves 94.65% average accuracy with 3.12% false positive and 2.24% false negative rates.