Dynamic Agricultural Pest Classification Using Enhanced SAO-CNN and Swarm Intelligence Optimization for UAVs

• Innovative model design: An enhanced SAO - CNN model integrating adaptive convolution, self - supervised learning, ConvLSTM, and bio - inspired algorithms is proposed. It improves recognition accuracy and robustness in complex agricultural settings. • Effective optimization strategies: Bio - inspired algorithms (e.g., genetic and particle swarm optimization) optimize UAV swarm intelligence, enhancing flight strategy, image acquisition, recognition efficiency, real - time performance, and coverage in large - scale farmland monitoring. • Temporal feature utilization: The ConvLSTM - based model combines constitutional layers and LSTM structure to extract temporal info from video frames, boosting recognition accuracy and timeliness in dynamic monitoring. • Superior experimental performance: The SAO - CNN model outperforms traditional and other deep learning models, achieving 91.2% classification accuracy, 0.89 recall, 0.88 F1 score, and 32 fps processing speed in agricultural pest image classification. The rapid advancement of agricultural modernization demands urgent solutions for accurate and real-time pest monitoring to enhance crop productivity. Traditional manual methods lack efficiency and fail to capture dynamic pest behaviors, while existing deep learning models struggle with robustness in complex environments. To address these challenges, this study proposes a novel Dynamic Agricultural Pest Classification System that integrates an enhanced Self-Activation Optimization Convolutional Neural Network (SAO-CNN) with bio-inspired swarm intelligence for UAVs. The SAO-CNN innovatively combines adaptive convolutional layers, self-supervised learning, and ConvLSTM to optimize spatial-temporal feature extraction, while swarm algorithms (ACO and PSO) enhance UAV path planning and task allocation. Key contributions include: (1) A hybrid SAO-CNN architecture that dynamically adjusts convolution kernels and leverages unlabeled data through self-supervised learning, improving adaptability to lighting and background variations. (2) A UAV swarm intelligence framework optimized via bio-inspired algorithms, reducing flight time by 29.2% and energy consumption by 32% compared to non-optimized systems. (3) Superior performance with 91.2% classification accuracy, 0.89 recall, and 32 FPS processing speed, outperforming state-of-the-art models (e.g., YOLO variants, ResNet, and ConvLSTM) in both static and dynamic scenarios. This work provides a robust solution for real-time pest monitoring, significantly advancing precision agriculture and sustainable crop management.