New AI model segments crop diseases with just a few images

Alongside the model, a high-quality benchmark dataset covering 101 pest and disease classes has been publicly released. Together, they offer a powerful and label-efficient solution for real-world plant health monitoring.

Pests and diseases cause 20–40% annual global crop loss (FAO), posing a direct threat to food security. Traditional detection relies heavily on manual observation, which is labor-intensive, subjective, and slow for large planting areas. While deep learning has accelerated automated crop health diagnostics, most progress has focused on image-level recognition rather than pixel-wise segmentation. Semantic segmentation—labeling every pixel in an image—can locate diseased areas with precision, yet requires extensive annotation. Field images further complicate the task due to lighting variability, background interference, and subtle symptom differences. Current few-shot approaches have rarely been applied in agriculture and often fail when lesions are small, scattered, or visually similar to surrounding tissue. These challenges highlight the need for a robust segmentation method that operates with limited labeled data.

A study (DOI: 10.1016/j.plaphe.2025.100121) published in Plant Phenomics on 30 September 2025 by Xijian Fan’s team, Nanjing Forestry University, presents a label-efficient few-shot semantic segmentation framework that enables accurate pixel-level detection of plant pests and diseases in real-world agricultural environments with minimal annotated samples.

In this study, the authors rigorously evaluated the proposed SegPPD-FS framework for few-shot semantic segmentation of plant pests and diseases using the mean intersection over union (mIoU) as the primary metric and foreground–background IoU (FB-IoU) as a supplementary indicator. They first benchmarked nine state-of-the-art FSS models (HDMNet, MSANet, MIANet, SegGPT, BAM, PFENet, DCP, PerSAM, and MGCL) on the SegPPD-101 dataset and selected HDMNet, which achieved the best mIoU, as the baseline to be improved. SegPPD-FS was then built by integrating two key modules—the similarity feature enhancement module (SFEM) and the hierarchical prior knowledge injection module (HPKIM)—to refine query features at different stages. All models were implemented in PyTorch and trained on a single NVIDIA GeForce RTX 4060Ti GPU, using ResNet50 or VGG16 backbones with PSPNet as a fixed feature extractor and meta-learning for the remaining components. Training was performed with AdamW over 150 epochs on the SegPPD-101 dataset, where 80 categories were used for training and 21 disjoint categories for combined validation/testing to assess cross-crop generalization under 1-, 2-, 4-, and 8-shot settings. Results show that SegPPD-FS consistently outperforms HDMNet and other FSS methods in mIoU and FB-IoU, achieving gains of up to 1.00% mIoU and 0.69% FB-IoU with ResNet50, and demonstrating particularly strong performance on objects of varying scales, although performance on small or rare classes remains more challenging. Qualitative comparisons confirm closer alignment with ground truth masks, with SFEM enhancing foreground discrimination and HPKIM effectively handling varying infestation severity, lighting conditions, and high background similarity. Ablation studies reveal performance drops when either SFEM or HPKIM is removed and show that an attention-based distillation loss improves learning, whereas an auxiliary loss and KL divergence-based variant can be detrimental. Despite slightly lower speed (5.14 FPS) than some competitors, SegPPD-FS offers roughly 10 percentage points higher accuracy and converges in about 60 epochs, indicating both efficient optimization and stable adaptation.

This research advances precision agriculture by reducing the heavy dependence on manual annotation and expert involvement. With the ability to learn from a handful of samples, SegPPD-FS offers an efficient tool for early warning diagnostics, digital field scouting, yield risk forecasting, and automated phenotyping. Its robust outputs may support integration into smart farming platforms, UAV-based surveillance, IoT crop monitoring systems, and large-scale disease mapping.

###

References

DOI

10.1016/j.plaphe.2025.100121

Original Source URl

https://doi.org/10.1016/j.plaphe.2025.100121

Funding information

This work was funded by the Key R&D Program of Jiangsu Province (BE2023369and BE2023352) and the Huai'an Science and Technology Plan Project (HAB202373).

About Plant Phenomics

Plant Phenomics is dedicated to publishing novel research that will advance all aspects of plant phenotyping from the cell to the plant population levels using innovative combinations of sensor systems and data analytics. Plant Phenomics aims also to connect phenomics to other science domains, such as genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics, and computer sciences. Plant Phenomics should thus contribute to advance plant sciences and agriculture/forestry/horticulture by addressing key scientific challenges in the area of plant phenomics.