Answers are grounded in this paper and cite their sources.
Dogwood ( Cornus florida ) is an economically valuable ornamental tree in North America, but leaf spot diseases reduce aesthetic quality and market value and are difficult to diagnose because symptoms overlap across pathogens. This study presents SpotChecker, an open-source, web-based decision-support application that prioritizes interpretable disease diagnosis and management guidance using a fine-tuned language model, supported by transformer-based image segmentation for quantitative disease severity estimation. A GPT-2 Medium language model was fine-tuned using LoRA on 3336 disease-specific question-answer (QA) pairs covering six major dogwood leaf spot diseases. The model generates the likely disease, causal pathogen, and concise management recommendations from user prompts, achieving 95% accuracy on a limited evaluation set of 60 manually constructed prompts. This design addresses the practical needs of growers, who often seek actionable guidance. To provide objective disease severity metrics that support and contextualize language-based recommendations, a transformer-based Mask2Former semantic segmentation framework was developed to delineate canopy and leaf spot regions from smartphone images. Standard segmentation architectures (U-Net, DeepLabV3+, and Feature Pyramid Network (FPN)) were first evaluated under a consistent patch-based training pipeline and showed limited performance in detecting small, sparse lesion regions. To address this limitation, a transformer-based Mask2Former framework was adopted and further extended with task-specific strategies. A dataset of 1100 annotated dogwood canopy images collected across the 2023–2024 seasons was used for training and evaluation. The Mask2Former framework was further extended using an inside-canopy two-class formulation, leaf-centered patch sampling, and an auxiliary binary cross-entropy (BCE)-Focal Tversky loss to address extreme lesion sparsity. Across controlled ablations (A2-A5), the final configuration improved validation leaf spot Intersection over Union (IoU) from 0.58 to 0.70, achieving test-time IoU values of 0.70 under patch-based evaluation and 0.60 under full-image sliding-window inference. Using segmentation outputs, SpotChecker computes continuous disease severity (%) as the proportion of predicted leaf spot pixels within the canopy. Across the test set, ground-truth severity ranged up to 3.46%, and SpotChecker produced severity estimates that closely matched ground truth across severity groups (mean predicted: 0.46%, 1.11%, and 2.06% for mild, moderate, and extreme vs 0.35%, 0.89%, and 2.11% ground truth, respectively), supporting objective tracking of disease progression and fungicide effectiveness. SpotChecker integrates these components in a Streamlit interface hosted on Hugging Face Spaces, allowing growers and practitioners to upload a smartphone image and receive both objective severity estimates and actionable, pathogen-aware guidance to support timely management decisions.
No takes yet. Share an insight, caveat, or question.
Rayamajhi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: