Review of the improved strategies for the visual detection of pest and disease using deep learning

Pests and diseases have seriously threatened the crop yield and quality, leading to economic loss. Timely and accurate detection of pests and diseases is often required for prevention and control in sustainable agriculture. Therefore, the detection of pests and diseases has been one of the most important hotspots in agricultural engineering. Conventional manual detection can also suffer from low efficiency due to its subjectivity and labor intensity. Fortunately, visual detection of pests and diseases can be developed using computer technology, particularly with its low cost and high speed. However, early image processing and conventional machine learning cannot fully meet the high accuracy in complex scenarios. Alternatively, deep learning has significantly improved the accuracy and adaptability of pest and disease detection in recent years. Furthermore, general architectures of deep learning are often required to better balance the different tasks of pest and disease detection. Complex targets and environments can also pose a great challenge to the limited techniques of deep learning. Existing reviews focus mostly on the basic frameworks of deep learning models. It is still lacking a systematic summary of improvements. Therefore, this study aims to review recent achievements in deep learning for the visual detection of pests and diseases. Improvement strategies were also categorized after evaluation. Among them, 1) the concepts and applications of deep learning were clarified for pest and disease detection. According to different tasks, visual detection of pests and diseases was divided into image classification, object detection, and image segmentation. 2) Commonly used improvements of deep learning were summarized from three aspects: training data, network models, and training strategies, with emphasis on the characteristics and scenarios of different approaches. Data augmentation, multimodal data fusion, and few-shot learning were widely used during data optimization. Specifically, data augmentation and few-shot learning were used to effectively alleviate data scarcity. Multimodal data fusion was added to effectively improve model representation. Network model improvement included attention mechanisms, model lightweighting, activation function improvements, multi-scale feature extraction and fusion, CNN-Transformer integration, residual connections, and ensemble deep learning. Among them, attention mechanisms, multi-scale feature extraction and fusion, and activation function improvements effectively enhanced feature extraction, thereby improving model performance. Model lightweighting significantly increased inference speed with high accuracy. CNN-Transformer integration was used to strengthen the complementarity of the architectures. Ensemble deep learning was suitable for scenarios where accuracy was required instead of speed. Transfer learning, pretraining, and fine-tuning effectively alleviated data scarcity for generalization from the perspective of optimization. Loss functions significantly enhanced optimization efficiency during training, according to different evaluation criteria. Key metrics were focused directly on alleviating the class imbalance. Additionally, sample selection, hyperparameter optimization, and optimizer enhancements were often used to optimize the training, such as loss descent strategies. The architectures were fully utilized to optimize training performance. Finally, the classical architectures and improvements were summarized for the future challenges in visual detection of pests and diseases using deep learning.