【摘要】 为应对传统方法在农业病虫害识别中效率低下、过度依赖专家知识的瓶颈，深度学习技术提供了新的技术途径。从“技术-场景-落地”的适应性视角出发，对该领域研究进行系统综述。通过构建“模型-数据-环境”分析框架，系统综述了基于CNN（卷积神经网络）与Transformer的分类检测模型及优化技术进展，剖析了公开数据集构建与标准化的现状与挑战，并深入探讨了模型泛化能力不足、小目标与遮挡物检测困难、硬件资源限制等制约技术落地的核心难题。在此基础上，提出了以“场景自适应”为核心的技术框架，强调需在数据、算法与部署三个层面进行协同创新，即构建贯通实验室与田间的渐进式数据集、发展具备在线学习与领域泛化能力的自适应模型及形成“云-边-端”协同的弹性计算架构，以期为推动深度学习在智慧农业中的深度融合与应用提供理论参考与实践指引。更多还原

【Abstract】 Deep learning offers a novel approach to overcoming the limitations of traditional methods for identifying agricultural pests and diseases, which are often inefficient and reliant on expert knowledge. From a "technologyscenario-application" adaptability perspective, a systematic review of research progress in this field is conducted. By constructing an analytical framework of "model-data-environment, " the advances in classification and detection models based on Convolutional Neural Network(CNN) and Transformer architectures and their optimization techniques are systematically reviewed. The current state and challenges in building and standardizing public datasets are analyzed, and the core obstacles hindering practical deployment, including insufficient model generalization, difficulties in detecting small and occluded objects, and hardware resource constraints are delved into. Furthermore, a "scenario-adaptive" technical framework as a core development pathway is proposed. This framework emphasizes synergistic innovation across three layers: constructing progressive datasets that bridge the laboratory-field gap at the data layer; developing self-adaptive models with online learning and domain generalization capabilities at the algorithm layer; and forming an elastic "cloud-edge-end" collaborative computing architecture at the deployment layer. It aims to provide theoretical reference and practical guidance for promoting the deep integration and application of deep learning in smart agriculture.更多还原