Skip to content

Research

Open questions, stated before the answers.

AgroExpert sits where multimodal AI, sustainable agriculture and environmental systems meet. This page sets out proposed research questions, how they could be tested, and what it would take to make any real-world claim. None of the evaluations below have been run yet.

Black-and-white archival photograph of scientists working at benches in a 1930s plant pathology laboratory.
CSIR Plant Pathology Laboratory, Canberra, c. 1932 - diagnosis has always depended on careful evidence. Photo: CSIRO Archives (opens in a new tab), CC BY 3.0.

Proposed directions, not results

There are no published papers, partners, experiments or findings from this project yet. Everything on this page describes what could be studied and how.

01Research questions

Five questions this project could help answer.

RQ1 · proposed

Can multimodal evidence improve crop-health assessment compared with image-only or text-only approaches?

Why it matters

Image-only models degrade sharply outside the conditions they were trained on. Context that a photo cannot show - crop stage, weather, recent activities - may resolve ambiguous cases.

Possible comparison

Image-only vs. text-only vs. image + plot context, on the same held-out cases.

Candidate metrics

  • Top-k diagnostic accuracy
  • Macro-F1 across causes
  • Calibration error

02Experimental comparisons

Ablations that isolate one design choice at a time.

Each comparison holds everything constant except one component, so its contribution can be measured on the same evaluation cases.
Proposed experimental comparisons
BaselineTreatmentTestsPrimary measures
LLM-onlyvs.Retrieval-grounded generationRQ2Faithfulness, citation correctness
Image-onlyvs.Multimodal evidenceRQ1Diagnostic accuracy, calibration
Generic advicevs.Context-aware adviceRQ3Expert-rated appropriateness
Unconstrained outputvs.Safety-verified outputRQ2, RQ4Unsafe and unnecessary recommendations

03Evaluation framework

Five dimensions of quality.

Accuracy alone is not enough. A useful system also needs honest confidence, traceable evidence, safe recommendations and the confidence of agricultural experts.

Dimension 1 of 5

Diagnostic quality

Is the possible diagnosis right - and is the confidence honest?

Candidate metrics

  • Top-k accuracy
  • Macro-F1
  • Expected calibration error
  • Abstention when evidence is weak

Where this stands today

Not yet measured: no labelled field-photo dataset exists.

SourcesEs et al. (2023) (opens in a new tab)Guo et al. (2017) (opens in a new tab)

04Environmental systems

Where decisions could meet environmental outcomes.

Solid links are established relationships, each with a source. Dashed links are this project’s hypotheses. The website does not measure or predict any of these outcomes.
Environmental systems contextA network linking farmer decisions to inputs and practices, and from there to soil health, crop productivity, water quality, emissions and farm income. The list beside the diagram describes every link.Farmer decisionFertilizer &pesticide inputsSoil & waterpracticesSoil healthCrop productivityWater qualityEmissionsFarm income
Established relationship (cited)Project hypothesis (to be tested)
Select a relationship

05Future environmental measures

What could be measured - with valid methods and data.

  • Fertilizer input intensity per unit area and per unit yield
  • Pesticide application frequency and hazard-weighted use
  • Nutrient-use efficiency where yield and input data are recorded
  • Water-quality indicators, where a valid monitoring partner exists
  • Greenhouse-gas estimates, only with an established method such as IPCC guidelines and adequate data

SourcesIntergovernmental Panel on Climate Change (2022) (opens in a new tab)FAO (web) (opens in a new tab)

06Before any claim

Prerequisites for real-world impact.

  • Agricultural expert review

    Agronomists and plant pathologists rating diagnoses and recommendations, blind to the system variant.

  • Representative datasets

    Field photos from the crops, regions, devices and conditions where the system would be used - not lab images.

  • Local validation

    Recommendations checked against local extension guidance, registrations and practices.

  • Field studies

    Real decisions and outcomes over seasons, with ethics review, consent and appropriate partners.

Farmers feed harvested wheat into a threshing machine as chaff piles up around them.
Threshing wheat after the winter harvest outside Lumbini, in the Terai of Nepal. Photo: Adam Jones (opens in a new tab), CC BY-SA 2.0.