💻 computer science

SAGE: Scalable Agentic Grounded Evaluation for Crop Disease Diagnosis

The SAGE framework addresses the scarcity of standardized plant disease data by introducing a large-scale, expert-validated image-symptom dataset and a scalable, training-free agentic pipeline that leverages grounded symptom knowledge to significantly improve the accuracy and explainability of crop disease diagnosis across diverse crops without requiring model retraining.

Original authors: Muhammad Arbab Arshad, Tirtho Roy, Yanben Shen, Dinakaran Elango, Shivani Chiranjeevi, Asheesh K. Singh, Baskar Ganapathysubramanian, Chinmay Hegde, Arti Singh, Soumik Sarkar

Published 2026-06-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Muhammad Arbab Arshad, Tirtho Roy, Yanben Shen, Dinakaran Elango, Shivani Chiranjeevi, Asheesh K. Singh, Baskar Ganapathysubramanian, Chinmay Hegde, Arti Singh, Soumik Sarkar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Blind" Doctor

Imagine a farmer has a sick plant. They take a picture and ask a computer, "What's wrong?"
Currently, most computer "doctors" are like blindfolded experts. They have memorized thousands of pictures of sick leaves, but if they see a new type of plant or a weird-looking disease, they often guess wrong. Worse, when they give an answer, they can't explain why. They just say, "It's Disease X," without showing their work.

The paper argues that to fix this, we need two things:

A massive, organized library of facts (not just pictures, but descriptions of symptoms with proof).
A "detective" AI that doesn't just guess, but investigates step-by-step.

The Solution: SAGE

The authors built a system called SAGE. Think of it as a super-intern medical student who has access to a giant, verified library and a magnifying glass.

1. The Giant Library (The Dataset)

The team created the largest plant disease library ever assembled.

The Scale: It covers 335 different crops (from soybeans to vanilla) and 1,251 different diseases. It contains nearly 840,000 images.
The Twist: Unlike other libraries that just have a picture and a label (e.g., "Tomato, Blight"), this library includes source-cited symptom descriptions.
- Analogy: Imagine a dictionary where every definition is backed by a direct quote from a specific encyclopedia page. If the library says "Blight causes brown spots," it also provides the exact sentence from a university website that proves it. This prevents the AI from "hallucinating" (making things up).

2. The Detective Agent (The AI)

Instead of a computer that looks at a photo and instantly spits out an answer, SAGE uses an autonomous agent (a digital detective).

How it works:
1. Observe: The agent looks at the sick plant and identifies the body part (e.g., "This is a leaf, not a stem").
2. Narrow Down: It checks its library. "Okay, this is a leaf. Which diseases affect leaves? Let's ignore the root diseases."
3. Investigate: The agent doesn't just guess. It has a "budget" (say, 8 chances) to look at reference images (pictures of known diseases).
4. Compare: It looks at the sick leaf, then looks at a reference image of "Frogeye Leaf Spot." It thinks: "Hmm, the reference has gray centers. My sick leaf has gray centers. Good match." Then it looks at another reference: "Rust has orange dots. My leaf doesn't have orange dots. Eliminate Rust."
5. Conclude: After checking a few references, it makes a final call and writes down a reasoning trace—a step-by-step story of how it solved the mystery.

The Results: Why It Matters

The team tested this detective on four crops: Soybeans, Corn, Tomatoes, and Mangoes.

The "Blind" Baseline: When the AI tried to guess without the library or the step-by-step investigation, it was often wrong.
The SAGE Boost: When the AI used the library and the detective method:
- Accuracy jumped significantly (by about 16 percentage points on average).
- It became much better at telling apart diseases that look very similar (like two different types of leaf spots).
The "Explainable" Win: The best part? You can read the detective's notes. You can see exactly which reference images it looked at and why it ruled out the other options. It's transparent, not a "black box."

Key Takeaways

No Retraining Needed: Because the AI uses a library of facts and reference images, you don't need to retrain the computer for every new crop. You just give it the new crop's library, and it can start diagnosing immediately.
Better than "One-Shot" Guessing: The paper shows that taking time to compare images one by one (like a detective) is much more accurate than trying to guess the answer in a single split second.
Trustworthy: Because every fact in the library is linked to a real web source with a quote, the system is grounded in reality, not just AI guesses.

In short: SAGE turns plant disease diagnosis from a "magic trick" into a transparent, step-by-step investigation backed by a massive, verified library of agricultural knowledge.

Technical Summary: SAGE (Scalable Agentic Grounded Evaluation) for Crop Disease Diagnosis

Problem Statement

Accurate plant disease diagnosis is critical for global food security, yet developing generalizable models remains hindered by two primary gaps:

Data Scarcity and Fragmentation: Existing datasets (e.g., PlantVillage, PlantDoc) are limited in crop diversity, often focus exclusively on leaves, and lack structured, verifiable symptom metadata. They do not support diagnosis across the vast spectrum of crops and pathogens found in real-world agriculture.
Lack of Explainability and Grounding: While deep learning and vision-language models (VLMs) have achieved high accuracy on controlled benchmarks, they often operate as "black boxes." They lack the ability to produce transparent, step-by-step diagnostic reasoning grounded in traceable, domain-specific knowledge sources. Furthermore, many existing agentic or chain-of-thought approaches require fine-tuning or operate on narrow crop sets.

Methodology

The paper introduces SAGE, a framework comprising a massive, curated dataset and a training-free agentic diagnostic pipeline.

1. The SAGE Dataset

The authors curated the largest plant disease image-symptom dataset to date, aggregating ~839,000 images covering 335 crops and 1,251 disease classes.

Multi-Organ Coverage: Unlike previous datasets focused on leaves, SAGE includes images of stems, roots, seeds, ears, and heads.
Source-Grounded Symptom Registry: A key innovation is the automated "Disease Registry Pipeline." Instead of relying on model-generated descriptions, the pipeline:
- Performs targeted web searches for each disease using authoritative sources (e.g., university extensions, CABI, APS compendia).
- Extracts structured fields (pathogen, affected organs, symptoms) with verbatim quotes from the source text.
- Reconciles naming variants and merges data across sources.
- Undergoes expert audit to validate field-level accuracy.
Anatomical Index: Images are tagged with the visible plant organ (leaf, stem, etc.), creating a mapping that allows the agent to narrow candidate diseases based on the organ present in a test image.

2. Agentic Diagnostic Pipeline

The core diagnostic system is an autonomous reasoning agent that operates without task-specific training (training-free). It utilizes a multi-turn loop to diagnose a test image:

Observation: The agent identifies the anatomical context (organ) and visible symptoms of the test image.
Candidate Narrowing: Using the Anatomical Index, it filters the candidate disease list to only those affecting the observed organ. It further prioritizes candidates using the Symptom Registry (if available).
Sequential Comparison: The agent operates under a reference budget ( $k$ ). It sequentially selects reference images from candidate classes, compares them to the test image, and reasons about visual matches and mismatches.
Trace Generation: The agent produces a fully explainable reasoning trace, documenting which references were viewed, the comparative analysis, and the final prediction. This ensures the decision process is auditable.

Key Contributions

SAGE Dataset: A multi-crop, multi-organ dataset of ~839K images with 1,251 disease classes, assembled from established benchmarks, expert collections, and community sources.
Source-First Disease Registry: A pipeline that generates structured symptom knowledge with per-field provenance (source URL and verbatim quote), eliminating hallucination risks associated with LLM-generated facts.
Training-Free Agentic Pipeline: A diagnostic system that leverages the registry and reference images to perform explainable, step-by-step reasoning without requiring model fine-tuning.
Systematic Evaluation: A comprehensive evaluation across four crops (Soybean, Corn, Tomato, Mango) with varying difficulty levels, testing different reference budgets and model tiers.

Results

The evaluation demonstrates that incorporating structured symptom knowledge and reference images significantly improves diagnostic accuracy:

Performance Gains: Incorporating the symptom knowledge base (KB) improved accuracy by an average of 16.2 percentage points across the four crops at a full reference budget ( $k=8$ $k = 8$ ).
- Soybean: KB alone (at $k=0$ ) added 14.1 pp.
- Corn: KB alone added 9.1 pp.
- Tomato: The full pipeline (KB + $k=8$ ) reached 76.1% accuracy, a 23.9 pp improvement over the no-KB baseline.
Reference Budget: Accuracy generally increased with the reference budget ( $k$ ), with the most significant gains observed in crops with higher disease class counts (Soybean, Corn).
Error Reduction: Confusion matrix analysis showed that the full pipeline reduced the over-prediction of visually similar diseases (e.g., distinguishing Sudden Death Syndrome from other soybean diseases) by leveraging KB symptoms and reference comparisons.
Model Independence: The framework benefits directly from improvements in foundation models (e.g., upgrading from Haiku to Opus) without retraining, as the agent relies on the model's inherent reasoning capabilities.

Significance and Claims

The paper claims that SAGE addresses the critical bottleneck of lacking large-scale, structured, and verifiable symptom knowledge for plant disease diagnosis. Its significance lies in:

Scalability: The framework can be extended to new crops without retraining, provided reference images and symptom knowledge are available.
Explainability: Unlike single-pass classification, the agentic approach provides a transparent reasoning trace, allowing users to audit why a diagnosis was made and which evidence was considered.
Grounded Reliability: By anchoring symptom descriptions to verbatim source quotes, the system mitigates hallucination and ensures diagnostic claims are traceable to authoritative domain knowledge.
Future-Proofing: The system is designed to leverage future advancements in foundation model capabilities automatically, making it a durable solution for scalable agricultural diagnostics.

The authors position this work as a step toward moving beyond opaque deep learning models toward diagnostic systems that are both accurate and grounded in verifiable, human-readable domain knowledge.