Computer vision and edge AI for crop health monitoring: a critical review of field validation and deployment evidence

Abstract Applications of computer vision include the diagnosis of plant diseases, the detection of pests and weeds, and the evaluation of abiotic stress. Although high benchmark performance does not necessarily translate into reliable field performance, this critical review evaluates the evidence on five questions regarding task formulation, validation, device-level implementation, robustness, and intervention effects. The existing knowledge base was updated by a targeted search and 10 selected articles were reviewed based on an extraction protocol. Due to the heterogeneity of the tasks, data sets, and validation settings, results are described and compared. The audited evidence illustrates that controlled-to-field performance losses can be substantial, the sensitivity of smartphone-based diagnosis to image-acquisition conditions, and the distinction between simulated resource estimates and measured device inference. Multimodal gains, uncertainty estimates, and visual explanations require separate validation under missing data, distribution shift, and agronomic use. We propose explicit evidence requirements for sensing affordability and deployment readiness, and distinguish these recommendations from validated study findings. Economic benefit and long-term intervention outcomes remain insufficiently established in the audited examples. The review supports locally validated diagnostic aids and supervised operational prototypes, while identifying the evidence needed for broader deployment claims. Its targeted selection does not constitute an exhaustive systematic review or a complete PRISMA-compliant database search.