Seedlabs
Source research

Recent advances in plant disease detection: challenges and opportunities

Muhammad Shafay, Taimur Hassan, Muhammad Owais, Irfan Hussain, Sajid Gul Khawaja, Lakmal D. Seneviratne et al. · 2025 · 116 citationsRead the paper

Plant diseases cause approximately 220 billion USD in annual agricultural losses, driving demand for automated detection systems. This systematic review analyzes deep learning approaches for plant disease detection using RGB and hyperspectral imaging, examining their evolution from classical image processing to modern neural architectures. We evaluate state-of-the-art models across 11 benchmark datasets, revealing significant performance gaps between laboratory conditions (95-99% accuracy) and field deployment (70-85% accuracy). Transformer-based architectures demonstrate superior robustness, with SWIN achieving 88% accuracy on real-world datasets compared to 53% for traditional CNNs. Our analysis identifies three critical deployment constraints: environmental variability sensitivity, economic barriers (500-2000 USD for RGB vs. 20,000-50,000 USD for hyperspectral systems), and interpretability requirements for farmer adoption. Case studies of successful platforms (Plantix with 10+ million users) highlight the importance of offline functionality and multilingual support. We establish evidence-based guidelines prioritizing deployment viability over laboratory optimization and identify key research directions including lightweight model design, cross-geographic generalization, and explainable multimodal fusion. This review provides a comprehensive framework for advancing plant disease detection from research prototypes to practical agricultural tools that can improve global food security.

What the full text shows

Working prototype
  • Significant performance gap exists between laboratory and field deployment accuracy. — 95-99% (lab) vs. 70-85% (field) (Across 11 benchmark datasets)
  • Hybrid architectures (SWIN) outperform traditional CNNs in real-world datasets. — 88.07% accuracy (SWIN) vs. 53% (traditional CNNs) (Cropped PlantDoc dataset)
  • Performance degrades significantly when moving from controlled to field environments. — ConvNeXt declined 27.71% (99.47% to 71.76%); ResNet50 declined 56.39% (98.18% to 41.79%) (PlantVillage to PlantDoc transition)
  • Adding spatial localization requirements reduces detection performance compared to simple classification. — 21.9 percentage point reduction (85.8% classification vs. 63.9% mAP) (PlantDoc dataset)
  • Few-shot learning (FSL) enables high accuracy with minimal labeled data. — ViT achieved 90.12% accuracy vs. ResNet50's 76.69% (5-way 5-shot classification)
  • Hyperspectral imaging (HSI) offers higher lab accuracy but faces extreme cost barriers. — 95-99% accuracy; cost 20,000-50,000 USD vs. 500-2,000 USD for RGB (Laboratory environments)
Setup
Evaluation across 11 benchmark datasets including PlantVillage, PlantDoc, FieldPlant, and Cassava Leaf Disease.
How
Systematic review and exploratory analysis of deep learning architectures (CNNs, ViT, SWIN) across multiple benchmark datasets. The study compares RGB and hyperspectral imaging modalities and evaluates classification vs. localization tasks.
Limitations
High sensitivity to environmental variability in field conditions · High cost and computational intensity of hyperspectral systems · Reduced interpretability of deep learning models compared to classical methods · Poor performance of detection (mAP) in complex field environments (FieldPlant)

Summarized by Gemma 4 from the open-access full text · 12 Oct 2026. Read it