#智能虫情测报灯

Agricultural pest surveillance is shifting from periodic manual inspection towards digitally mediated observation, combining automated traps, computer vision, environmental sensors and statistical or machine-learning forecasts. The technological literature reports increasingly sophisticated species recognition, remote counting and near-real-time data transmission. Whether these capabilities generate reliable estimates of pest pressure or improve integrated pest management, however, depends on ecological sampling, field validation and decisions taken after an alert. This critical narrative review evaluates the measurement-to-decision pathway rather than treating classification accuracy as evidence of agronomic benefit. Peer-reviewed research and relevant methodological literature were identified through accessible scholarly sources, with thematic emphasis on camera-equipped pheromone, light and adhesive traps; image-recognition architectures; optical, acoustic and environmental sensing; predictive modelling; and operational evaluation. The synthesis indicates that instrumented traps and image-processing methods can reduce some observation burdens and produce useful within-system temporal information. Their reported performance is nevertheless strongly conditioned by trap selectivity, target species, illumination, specimen overlap, annotation quality, seasonal distribution shifts and the choice of validation dataset. Forecasting approaches can exploit temperature, phenology and historical trap catches, but retrospective fit does not establish transferable warning value. Ecological inference also requires distinguishing insect activity, trap catches, abundance, crop injury and economic risk. The most consequential weaknesses are limited independent multi-site prospective testing, incomplete uncertainty reporting, insufficient comparison with manual scouting, and few controlled evaluations of management consequences. A deployment-oriented framework is proposed in which measurement validity, calibration, forecast skill, decision thresholds, costs and non-target effects are appraised as connected requirements. Progress will depend less on marginal improvements in benchmark detection scores than on prospective, locally calibrated studies that measure timeliness, intervention quality, labour, pesticide use, yield and biodiversity outcomes.

综述AI（CNN、ViT、无人机遥感、物联网）用于作物病虫害早期检测并融入种植园风险管理。

利用智慧农业站数据评估油桃园蚜虫风险，发现10月为高风险期。

山东青岛市城阳区夏庄街道上山色峪村中华蜜蜂产业园是山东省首家规模化智慧蜂场，智慧蜂箱搭载物联网传感器，箱内温湿度、蜂群活跃度、蜜蜂出勤等数据实时回传后台终端，蜂农用手机即可掌握蜂群状态。产蜜量同比提升20%，蜂蜜品质全程可追溯。

设计并评估温室四参量LoRaWAN监测节点，测试多节点丢包与信号性能。

提出基于ESP32与CNN的智能农业系统，实现环境监测与叶片病害实时检测。

近日，中国农业科学院农业信息研究所科学数据研究室科研团队提出一种层次知识引导的枸杞害虫识别框架，有效解决真实农业场景下样本稀缺与类间混淆的双重挑战。相关研究成果发表在《人工智能的工程应用（Engineering Applications of Artificial Intelligence）》上。科研团队依托长文本-图像检索模型，把专家撰写的长文本描述解析为粗、中、细三个语义层次，并与多尺度视觉特征实现精准跨模态对齐；同时设计知识引导的去混淆模块，通过构建文本和视觉双模态相似度图，明显降低容易混淆类别之间的预测误差。该研究为枸杞等特色作物的虫害智能监测与精准防控提供新的技术路径。

研发基于无人机与轻量YOLO的水稻病虫害智能检测系统Agri-IntelliDet。