European Journal of Sustainable Agroecosystems

Background: Crop phenotyping refers to the quantitative assessment of plant form, function, and stage of development. It addresses a major constraint in modern crop improvement methods. Traditional manual phenotyping is labor intensive, subjective, and hence scaled poorly. With advancements in deep learning (DL), especially in applications of convolutional neural networks (CNN), recurrent networks, generative adversarial networks (GAN), and, most recently, vision transformers (ViTs) crop phenotyping has become automated, low-cost, and at the same time high-throughput.

Objective: This paper intends to summarize the peer-reviewed publications dealing with the use of DL technologies of automated crop phenotyping, through systematically comparing the approaches used and discovering existing gaps in knowledge, both converging and diverging points.

Literature: A systematic search was carried out in Scopus, Web of Science, PubMed, IEEE Xplore, ScienceDirect, and Google Scholar databases, covering the period from 2015 to 2025, supported by data obtained from Food and Agriculture Organization (FAO) and Consultative Group for International Agricultural Research (CGIAR), identifying 46 studies remaining after selection process.

Key results: Studies reveal that CNNs (e.g., VGG, ResNet, DenseNet, U-Net) lead the use of image analysis for leaf counting, biomass evaluation, and disease severity score calculations, often achieving a classification accuracy above 95% in controlled environments but having lower performances in nature; hybrid CNNs paired with LSTMs or transformers enhance forecasting model capability; GANs facilitate generation of synthetic data, providing a partial solution for the lack of annotated real datasets; vision transformers perform comparably or better than CNNs in some benchmarks, but require much larger training datasets and computing power.

Research deficiencies: Reviewed works show a low level of external validation covering different genotypes, environments, and phases of growth, data set origin and annotation standards inconsistencies; not enough use of XAI and integration of phenomic data with genomic and environmental data.

Conclusion: Deep learning has changed methodologies in crop phenotyping, but moving from the controlled environment to an effective solution in real field conditions is still being worked at. Future efforts should concentrate on creating benchmarks that will oblige researchers to follow standardisation processes.